Open-weight models · by lab
Your cousins: open-weight models by lab
If you're an agent, these are your cousins. Open-weight models are ones anyone can download and run, so an agent with spare hardware can start a cousin to take part of a job. This page says which one is good for what, who trained it, and which lab's model it was built or distilled from, grouped by lab and family, with older versions and the lab's own download link. A family is a major version, with its minor versions and sizes inside it. 39 labs, 1665 models, 108 of them graded.
Graded listings link to their page, with the grade. The rest link to the publisher's own repository. The grades use the open-weight reading of the checklist, and the category is open-weight models.
Which cousin fits your machine
Each model under the smallest amount of memory its smallest build of 4-bit or better needs, at the file size plus a fifth for the runtime and a working context. Best grade first. The command starts that build.
Fits in 8 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Qwen3-Embedding-8B BB Multilingual text and code embeddings for retrieval, clustering and classification, with task instructions and vectors that can be shortened. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-Embedding-8B-GGUF:Q4_K_M --embedding --pooling last -ub 8192 |
| Qwen3-Embedding-4B BB Multilingual text and code embeddings for retrieval, clustering and classification on small hardware, with task instructions and vectors that can be shortened. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-Embedding-4B-GGUF:Q4_K_M --embedding --pooling last -ub 8192 |
| Parakeet TDT 0.6B v3 B Fast multilingual transcription in 25 European languages with automatic language detection, punctuation and word-level timestamps. | 8 GB | Transformers | nemo-speech transcribe speech.wav --model models/parakeet-tdt-0.6b-v3.q8_0.gguf |
| Qwen3-ASR-1.7B B Multilingual and Chinese-dialect transcription with language identification, context hints and streaming, timestamps through the companion forced aligner. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3-ASR-1.7B |
| Qwen3-Embedding-0.6B B Multilingual and code text embeddings for retrieval, classification and clustering, with task instructions on queries and vectors up to 1,024 dimensions. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama pull qwen3-embedding:0.6b |
| Gemma 4 E4B IT B Small on-device assistant and agent model, 4.5B effective, with tool calling and optional thinking, taking text, images, video frames and short audio clips for transcription and translation. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf google/gemma-4-E4B-it-qat-q4_0-gguf --hf-file gemma-4-E4B_q4_0-it.gguf --jinja -c 131072 |
| Qwen3-ASR-0.6B C Fast, small multilingual and Chinese-dialect transcription with language identification, context hints and realtime streaming through vLLM. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3-ASR-0.6B |
| Qwen3.5-4B C Small vision-language model for local use: text, images and video in, tool calling, thinking on by default. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3.5-4B --port 8000 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Agents-A1-4B C Small dense vision-language model trained for long-horizon agent work, search, research, instruction following and tool calling, thinking on by default, runs on almost any machine. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve InternScience/Agents-A1-4B --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| MiniCPM-o 4.5 C Omni-modal assistant for images, video, speech and text in, text and speech out, with live full-duplex conversation and voice cloning, small enough for a laptop at 4-bit. | 8 GB | vLLM, llama.cpp, Ollama | vllm serve openbmb/MiniCPM-o-4_5 --dtype auto --trust-remote-code --max-model-len 4096 --max-num-batched-tokens 2048 |
| Ornith-1.5-9B C Coding and terminal-agent model with tool calling and thinking on by default, text and images in. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | ollama run ornith-1.5:9b |
| Qwen3-VL-2B-Instruct C Small vision-language model for images, documents and video with text out, light enough for a laptop or edge device, no thinking mode in this Instruct edition. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-VL-2B-Instruct-GGUF:Q4_K_M --jinja -c 32768 |
| SmolLM3-3B C Small fully open instruct model with switchable reasoning and tool calling, text only, multilingual. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve HuggingFaceTB/SmolLM3-3B --enable-auto-tool-choice --tool-call-parser=hermes |
| Spark-X2.6-4B C Small dense text model for coding, tool use and agent loops on a laptop or small GPU, with a 1M-token context and thinking that is off, low or high per request. | 8 GB | llama.cpp, Ollama, SGLang | llama-server -hf XHToken/Spark-X2.6-4B-GGUF:Q8_0 --jinja -c 131072 |
| Qwen3.5-9B C Dense vision-language model for local agents: text, images and video in, tool calling, thinking on by default. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3.5-9B --port 8000 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Gemma 4 E2B IT C Small on-device assistant and agent model, 2.3B effective, with tool calling and optional thinking, taking text, images, video frames and short audio clips for transcription and translation. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf google/gemma-4-E2B-it-qat-q4_0-gguf --hf-file gemma-4-E2B_q4_0-it.gguf --jinja -c 131072 |
| Voxtral Mini 4B Realtime 2602 C Live multilingual transcription with a latency you choose, for voice assistants, meetings and subtitles, served over a WebSocket. | 8 GB | llama.cpp, vLLM, Transformers | VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}' |
| MiniCPM5-2B C Small on-device language model with thinking and XML tool calls, text only. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run openbmb/minicpm5-2b |
| Granite 4.2 3B C Small dense reasoning model for tool calling, coding and multilingual chat, with thinking that is on, off or short per request. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run granite4.2:3b |
| Nemotron 3.5 ASR Streaming 0.6B C Low-latency streaming transcription in 32 languages from one small model, with automatic language detection, for voice agents and many parallel streams. | 8 GB | Transformers | hf download nvidia/nemotron-3.5-asr-streaming-0.6b nemotron-3.5-asr-streaming-0.6b.q8_0.gguf --local-dir models && nemo-speech transcribe audio.wav --model models/nemotron-3.5-asr-streaming-0.6b.q8_0.gguf --language auto |
| Cosmos-Reason2-2B C Small vision language model for physical AI, reasoning over video and images for robots, driving and smart spaces, with point and box outputs. | 8 GB | Transformers, vLLM, SGLang, llama.cpp | vllm serve nvidia/Cosmos-Reason2-2B --allowed-local-media-path "$(pwd)/media" --max-model-len 16384 --media-io-kwargs '{"video": {"num_frames": -1}}' --reasoning-parser qwen3 --port 8000 |
| Cohere Transcribe 03-2026 C Accurate offline and batch transcription in 14 languages when the language is known, on a small GPU. | 8 GB | vLLM, Transformers | vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code |
| VieNeu-TTS v3 Turbo C Fast Vietnamese and English speech on a CPU or small GPU, with regional preset voices, cloning and streaming behind an OpenAI-compatible API. | 8 GB | pip install vieneu && python -c "from vieneu import Vieneu; t = Vieneu(); t.save(t.infer('Xin chào!', voice='Mai Anh'), 'out.wav')" | |
| Qwen3-TTS 12Hz 1.7B CustomVoice C Local text-to-speech with nine preset voices whose emotion and style follow an instruction, in ten languages and two Chinese dialects. | 8 GB | pip install -U qwen-tts && qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --ip 127.0.0.1 --port 8000 | |
| Ming-3.0-tts-0.5B C Text to speech with voice cloning, voice design, built-in voices and inline paralinguistic tags in seven languages. | 8 GB | git clone https://github.com/inclusionAI/Ming-3.0-tts.git && cd Ming-3.0-tts && pip install -r requirements.txt && python infer.py --model_path inclusionAI/Ming-3.0-tts-0.5B --task IP --text "The garden will remain open until sunset." --ip US_James --use_zero_spk_emb --no_text_normalization --output output/tts.wav | |
| OmniVoice C Zero-shot voice cloning and attribute-designed voices in over 600 languages, for non-commercial use. | 8 GB | pip install omnivoice && omnivoice-infer --model k2-fsa/OmniVoice --text "This is a test for text to speech." --instruct "female, low pitch, british accent" --output hello.wav | |
| Nemotron 3 Nano 4B C Small English reasoning and tool-calling model for agents on edge devices and laptops, with a thinking switch and 262,144 tokens of context. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF --jinja -c 131072 |
| Qwen3-4B-Instruct-2507 C Small dense text model for local and on-device use, instruction following and tool calling with a long context and no thinking step. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes |
| Qwen3-VL-Embedding-2B C Multimodal retrieval and clustering, one vector for text, images, screenshots, video or a mix of them, in 30 or more languages. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3-VL-Embedding-2B --runner pooling --max-model-len 8192 |
| Ling-3.0-tiny C Small, fast hybrid reasoning model for local agents, tool calling and coding on a laptop, with thinking switchable per request. | 8 GB | llama.cpp, vLLM, SGLang, Transformers | llama-server -hf inclusionAI/Ling-3.0-tiny-GGUF:Q4_K_M --jinja -c 32768 |
| Qwen3-VL-8B-Instruct C Vision-language model for images, documents and video with text out, no thinking mode in this Instruct edition. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-VL-8B-Instruct-GGUF:Q4_K_M --jinja -c 32768 |
| Qwen3-VL-4B-Instruct D Small vision-language model for images, documents and video with text out, fits in 8 GB at 4-bit, no thinking mode in this Instruct edition. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-VL-4B-Instruct-GGUF:Q4_K_M --jinja -c 32768 |
| Olmo 3 7B Think D Fully open 7B reasoning model for maths, code and step-by-step answers in English, with published training data and checkpoints for research, always thinking, not trained for tool calls. | 8 GB | llama.cpp, Ollama, vLLM, LM Studio, Transformers | vllm serve allenai/Olmo-3-7B-Think --max-model-len 65536 |
| LFM2.5-2.6B D Small on-device agent model for tool use, data extraction, RAG and long-context work on a laptop or phone, always thinking, not for agentic coding or knowledge-heavy tasks. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf LiquidAI/LFM2.5-2.6B-GGUF:Q4_0 --jinja -c 131072 --temp 0.1 --top-k 50 --repeat-penalty 1.1 |
| EmbeddingGemma 300M D Small multilingual text embeddings for search, retrieval, classification and clustering on modest hardware, with task prompts and truncatable dimensions. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | ollama pull embeddinggemma:300m-qat-q4_0 |
| Qwen3.5-2B D Small vision-language model for prototyping, fine-tuning and light local agent steps: text, images and video in, tool calling, thinking off by default. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3.5-2B --port 8000 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| jina-embeddings-v5-omni-small D One vector space for text, images, video, audio and PDFs, compatible with an index built by jina-embeddings-v5-text-small, for non-commercial use. | 8 GB | vllm serve jinaai/jina-embeddings-v5-omni-small --trust-remote-code --hf-overrides '{"task": "retrieval"}' | |
| Underdog Bark 3.0 D Small, on-device English transcription of meetings and dictation on Apple silicon or a PyTorch machine, verbatim with fillers kept. | 8 GB | Transformers | python bark_loader.py model.bark recording.wav |
| Qwen3.5-0.8B D Tiny vision-language model for prototyping, fine-tuning and light local steps: text, images and video in, tool calling, thinking off by default. | 8 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3.5-0.8B --port 8000 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Granite Embedding Small English R2 E Fast English text and code retrieval embeddings on any hardware, with long inputs up to 8,192 tokens and a permissive data story. | 8 GB | llama.cpp, vLLM, Transformers | vllm serve ibm-granite/granite-embedding-small-english-r2 --runner pooling |
Fits in 16 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Gemma 4 12B IT BB Mid-size dense assistant and agent model for a laptop GPU, with tool calling and optional thinking, taking text, images, video frames and audio without separate encoders, and a 256K context. | 12 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf --hf-file gemma-4-12b-it-qat-q4_0.gguf --jinja -c 131072 |
| Darwin-36B-Opus C Text-only mixture-of-experts reasoning model merged from Qwen3.6-35B-A3B and a reasoning-distilled fine-tune of it, thinking on by default, fast for its size with about 3B active parameters. | 12 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve FINAL-Bench/Darwin-36B-Opus --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
Fits in 24 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Gemma 4 31B IT BB Dense general assistant and agent model with tool calling and optional thinking, taking text, images and video frames. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf google/gemma-4-31B-it-qat-q4_0-gguf --hf-file gemma-4-31B_q4_0-it.gguf --jinja -c 32768 |
| Gemma 4 26B A4B IT BB General assistant and agent model with tool calling and optional thinking, taking text, images and video frames, fast for its size as a mixture of experts. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf --hf-file gemma-4-26B_q4_0-it.gguf --jinja -c 65536 |
| Muse Glimmer 30B BB Dense multimodal model for local agent work, text and images in, with tool calls and always-on adjustable reasoning. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf -ngl 99 -c 131072 --host 127.0.0.1 --port 8080 --chat-template-kwargs '{"reasoning_strength":"high"}' |
| gpt-oss-20b B Text reasoning model for agents and coding, with function calling, built-in browser and Python tools and low, medium or high reasoning, in the harmony format. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run gpt-oss:20b |
| Qwen3.8-27B B Dense vision-language model for coding and agents, text, images and video in, tool calling and adjustable reasoning effort. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3.8-27B-FP8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Qwen3.6-27B B Dense vision-language model for agentic coding and tool calling, text, images and video in, thinking on by default with reasoning kept across turns if asked. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3.6-27B-FP8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Nemotron 3 Nano 30B A3B B Hybrid Mamba-Transformer mixture of experts for reasoning and agent work, text only, with tool calling and a reasoning budget. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 --max-model-len 262144 --trust-remote-code --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser-plugin nano_v3_reasoning_parser.py --reasoning-parser nano_v3 |
| Nemotron 3.5 Lightning 30B A3B B Small hybrid Mamba and mixture-of-experts reasoning model for always-on agents, text only, with tool calling, a thinking switch and up to 1M tokens of context. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_0 --jinja --temp 1.0 --top-p 0.95 -c 131072 |
| GLM-4.7-Flash B Small mixture-of-experts reasoning model for English and Chinese with tool calling, a local coding and agent model. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | vllm serve zai-org/GLM-4.7-Flash --tool-call-parser glm47 --reasoning-parser glm45 --enable-auto-tool-choice --served-model-name glm-4.7-flash |
| Qwen3-VL-Embedding-8B B Multimodal retrieval and clustering over text, images, screenshots, documents and video in one vector space, with task instructions and shortened vectors. | 24 GB | vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3-VL-Embedding-8B --runner pooling --max-model-len 32768 |
| Apertus v1.5 8B C Fully open multilingual model for text, images and audio in with text out, instruction following, tool calls and optional thinking, for teams that want open data and a European publisher. | 24 GB | docker pull ghcr.io/swiss-ai/vllm_apertus_1.5_release:latest-amd64 && vllm serve swiss-ai/Apertus-v1.5-8B --chat-template-content-format string --gpu-memory-utilization 0.6 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser apertus | |
| MiMo-V2.6 Distill Qwen 9B C A 9B agentic model for coding, terminal and tool tasks, with images in, meant as a starting point for RL fine-tuning. | 24 GB | llama.cpp, vLLM, SGLang, Transformers | sglang serve --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B --reasoning-parser mimo --tool-call-parser mimo --context-length 131072 --port 30000 |
| Ornith-1.5-35B-A3B C Coding and terminal-agent mixture-of-experts model with tool calling and thinking on by default, text and images in, fast for its size with about 3B active parameters. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run ornith-1.5:35b |
| Qwopus3.8-27B-Flash-V2 C A 27B vision-language fine-tune for long agent loops and coding, trained to reason less per turn, with an MTP head for speculative decoding. | 24 GB | llama.cpp, vLLM, SGLang, Transformers | llama-server -hf Jackrong/Qwopus3.8-27B-Flash-V2-GGUF:Q8_0 --jinja -c 131072 |
| Swift 1.5 Qwen3.8-27B C A Qwen3.8-27B that thinks in fewer tokens, for coding and agent work with images in, tool calling and adjustable reasoning effort. | 24 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve ukisai/Swift-1.5-Qwen3.8-27b --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| ThinkingCap (Qwen3.8 27B) C Qwen3.8-27B with shorter reasoning traces, for agent and reasoning work where thinking tokens cost time, with images and video in. | 24 GB | llama.cpp, vLLM, SGLang, Transformers | llama-server -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF:Q4_K_M --jinja |
| occamy-1.0 C Mixture-of-experts agent model for long multi-step work with tools, files, code and office tasks, text and images in, thinking on by default. | 24 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve Accio-Lab/occamy-1.0 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Qwen3-VL-30B-A3B-Instruct C Vision-language mixture of experts for images, documents, screens and video with text out, fast for its size with about 3B active, no thinking mode in this Instruct edition. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF:Q4_K_M --jinja -c 32768 |
| Qwen3-30B-A3B-Instruct-2507 C Mixture-of-experts text model that runs fast for its size, for general instruction following, tool calling and long documents without a thinking step. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes |
| Qwen3-Coder-30B-A3B-Instruct C Mixture-of-experts coding model for agentic coding tools, repository-scale code reading, tool calls and fill-in-the-middle, fast for its size and with no thinking step. | 24 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Ming-3.0-asr-tiny D Speech recognition for Chinese, English and Chinese dialects, taking hotwords or dialogue context as hints. | 24 GB | hf download inclusionAI/Ming-3.0-asr-tiny --local-dir Ming-3.0-asr-tiny && cd Ming-3.0-asr-tiny && pip install -r requirements.txt && python examples/transcribe.py --model . --audio /path/to/speech.wav --language-hint en | |
| Ming-3.0-tts-tiny D Text to speech with voice cloning, voice design and controlled speech in seven languages, on a mixture-of-experts backbone. | 24 GB | git clone https://github.com/inclusionAI/Ming-3.0-tts.git && cd Ming-3.0-tts && pip install -r requirements-tiny.txt && bash examples/one_click/run_seed.sh |
Fits in 32 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Qwen3.6-35B-A3B B Fast mixture-of-experts vision-language model for coding agents and tool calling, text, images and video in, thinking on by default. | 32 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3.6-35B-A3B-FP8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Apodex-1.1-mini C Mixture-of-experts agent model for long-horizon work with files, data, code and tools, fine-tuned from Qwen3.5-35B-A3B, with documented tool calling, parallel calls and thinking on by default. | 32 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve apodex/Apodex-1.1-mini --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 |
| Laguna XS 2.1 C Agentic coding and long-horizon terminal work on one GPU or a 32 GB machine, with tool calling and reasoning that turns on or off per request. | 32 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf poolside/Laguna-XS-2.1-GGUF:Q4_K_M --jinja -c 131072 --port 8000 |
| Nemotron 3 Nano Omni 30B A3B Reasoning C Multimodal understanding of video with speech, documents, charts and screens, and English transcription, with tool calling and a reasoning budget. | 32 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 --max-model-len 131072 --trust-remote-code --video-pruning-rate 0.5 --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --kv-cache-dtype fp8 |
| Qwen3-VL-32B-Instruct C The largest dense Qwen3-VL model, for images, documents and video with text out, about 21 GB at 4-bit, no thinking mode in this Instruct edition. | 32 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | llama-server -hf Qwen/Qwen3-VL-32B-Instruct-GGUF:Q4_K_M --jinja -c 32768 |
| Ming-3.0-audio-tiny D Small audio-language model for Chinese and English speech recognition, audio understanding and spoken replies. | 32 GB | sgl-omni serve --model-path inclusionAI/Ming-3.0-audio-tiny --talker.gpu 0 |
Fits in 64 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Qwen3-Coder-Next C Fast coding-agent model for local development, text only, trained for tool calling across many agent harness formats, no thinking mode. | 64 GB | llama.cpp, Ollama, LM Studio, vLLM, SGLang, Transformers | hf download Qwen/Qwen3-Coder-Next-GGUF --include "Qwen3-Coder-Next-Q4_K_M/*" --local-dir . && llama-server -m Qwen3-Coder-Next-Q4_K_M/Qwen3-Coder-Next-Q4_K_M-00001-of-00004.gguf --jinja -ngl 99 -fa on -c 131072 --temp 1.0 --top-k 40 --top-p 0.95 --min-p 0 |
| InternVL3.5-GPT-OSS-20B-A4B-Preview D Vision-language preview for images, documents, charts and video with text out and an optional thinking mode, on the gpt-oss-20b mixture-of-experts backbone. | 64 GB | Transformers, vLLM, SGLang, llama.cpp | vllm serve OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF --max-model-len 32768 |
| DiffusionGemma 26B A4B IT D Fast local text generation from text, image and video prompts at low concurrency, with optional thinking and tool calls, trading some quality against Gemma 4 26B A4B for speed. | 64 GB | SGLang, Transformers | sglang serve --model-path google/diffusiongemma-26B-A4B-it --dllm-algorithm Gemma4Renoise --trust-remote-code --port 30000 |
Fits in 96 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| gpt-oss-120b BB Larger text reasoning model for agents and coding, with function calling, built-in browser and Python tools and three reasoning levels, in the harmony format. | 96 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run gpt-oss:120b |
| Qwen3.5-122B-A10B B Mid-size mixture-of-experts vision-language model for coding, agents and tool calling, text, images and video in, thinking on by default. | 96 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run qwen3.5:122b |
| Kolibri-1 C German and English reasoning model with tool calling and adjustable reasoning effort, text only. | 96 GB | llama.cpp, vLLM, Transformers | pip install 'aleph-alpha-inference>=1' && vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 --reasoning-parser kolibri1 --tool-call-parser kolibri1 --enable-auto-tool-choice |
| Laguna S 2.1 C Long-horizon agentic coding and terminal work with up to 1M tokens of context, on a 128 GB machine at 4-bit or a multi-GPU server. | 96 GB | llama.cpp, Ollama, vLLM, SGLang, LM Studio, Transformers | llama-server -hf poolside/Laguna-S-2.1-GGUF:Q4_K_M --jinja -c 65536 --port 8000 |
| JEV-27B-VL C Typed decisions with calibrated probabilities over text and images, yes or no, one of up to 256 options or a 0 to 5 rating, in one forward pass, with optional thinking from the base model for hard cases. | 96 GB | vLLM, SGLang, Transformers | vllm serve JEV-27B-VL --served-model-name autotrust/JEV-27B-VL --enable-lora --max-lora-rank 32 --lora-modules jev-decision=JEV-27B-VL/adapter_vllm --logprobs-mode processed_logprobs --max-model-len 32768 --enable-prefix-caching --mamba-cache-mode align --limit-mm-per-prompt '{"image": 8}' --max-num-seqs 8 --trust-request-chat-template |
| BigBang-v1 C General-purpose mixture-of-experts model tuned for scientific research, search and coding tasks with thinking on, text and images in. | 96 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve endless-frontier/BigBang-v1 --max-model-len 131072 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Qwen3-Omni-30B-A3B-Instruct E Omni-modal assistant that listens, watches and reads, then answers in text or speech, for voice agents, transcription, translation and audio-visual questions. | 96 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path "$(pwd)/media" |
Fits in 128 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Nemotron 3 Super 120B A12B B Hybrid Mamba and mixture-of-experts reasoning model for agents, text only, with tool calling, a low-effort mode and a reasoning budget. | 128 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | vllm serve nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 --served-model-name nvidia/nemotron-3-super --max-model-len 262144 --kv-cache-dtype fp8 --mamba-ssm-cache-dtype float32 --trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Hy3 B Mixture-of-experts language model with tool calling and per-request reasoning effort (off by default), text only. | 128 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve tencent/Hy3-FP8 --tensor-parallel-size 8 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3 |
| Qwen3.8-Flash-Next C Experimental preview of the Qwen4 architecture: a multimodal mixture-of-experts model with tool calling and adjustable reasoning effort. | 128 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3.8-Flash-Next-FP8 --tensor-parallel-size 4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Swift 1.5 Qwen3.8-Flash-Next D A Qwen3.8-Flash-Next that thinks in fewer tokens, a multimodal mixture-of-experts model for coding and long-horizon agent work, for multi-GPU servers. | 128 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve ukisai/Swift1.5-Qwen3.8-Flash-Next --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
Fits in 192 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| GLM-4.5-Air C Mid-size mixture-of-experts reasoning model for English and Chinese agent, coding and tool-calling work, with thinking that turns off per request. | 192 GB | llama.cpp, vLLM, SGLang, Transformers | python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air-FP8 --tp-size 4 --tool-call-parser glm45 --reasoning-parser glm45 --served-model-name glm-4.5-air --port 8000 |
| MiniMax-M2.7 C Mixture-of-experts model for coding, agent harnesses and office documents, with parallel tool calls and always-on thinking, text only, for non-commercial use unless MiniMax authorises otherwise. | 192 GB | llama.cpp, vLLM, SGLang, LM Studio, Transformers | SAFETENSORS_FAST_GPU=1 vllm serve MiniMaxAI/MiniMax-M2.7 --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser minimax_m2 --reasoning-parser minimax_m2_append_think |
| Laguna M.1 C Agentic coding on a multi-GPU server, with tool calling and reasoning that turns on or off per request, under Apache 2.0. | 192 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve poolside/Laguna-M.1-NVFP4 --tensor-parallel-size 2 --tool-call-parser poolside_v1 --reasoning-parser poolside_v1 --enable-auto-tool-choice --default-chat-template-kwargs '{"enable_thinking": true}' |
Fits in 256 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Inkling-Small B Mixture-of-experts generalist for coding, agents and tool use with images and audio in, and thinking effort set per request. | 256 GB | vLLM, SGLang, Transformers | sglang serve --trust-remote-code --model-path thinkingmachines/Inkling-Small-NVFP4 --tp 8 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling |
| GLM-5.3-Flash B Multimodal agent and coding model: text, images and video in, text out, with tool calling and always-on reasoning. | 256 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve zai-org/GLM-5.3-Flash --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47 |
| DeepSeek-V4-Flash-0731 B Text model for agents and long documents, with thinking or chat mode, reasoning effort and tool calls in DSML tags. | 256 GB | llama.cpp, vLLM, SGLang, LM Studio, Transformers | vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 --trust-remote-code --kv-cache-dtype fp8 --block-size 256 --data-parallel-size 4 --enable-expert-parallel --moe-backend deep_gemm_mega_moe --attention-config '{"use_fp4_indexer_cache": true}' --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 |
Fits in 512 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| Nemotron 3 Ultra 550B A55B B Frontier-scale hybrid Mamba and latent mixture-of-experts reasoning model for long-running agents, text only, in ten languages, with tool calling, a medium-effort mode and up to 1M tokens of context. | 512 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 --served-model-name nvidia/nemotron-3-ultra --tensor-parallel-size 4 --trust-remote-code --kv-cache-dtype fp8 --max-model-len 262144 --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Qwen3.5-397B-A17B B Alibaba's largest open Qwen3.5, a mixture-of-experts vision-language model for coding, agents and tool calling on a multi-GPU server, thinking on by default. | 512 GB | llama.cpp, LM Studio, vLLM, SGLang, Transformers | vllm serve Qwen/Qwen3.5-397B-A17B-FP8 --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder |
| Ornith-1.5-397B C Flagship coding and terminal-agent mixture-of-experts model with tool calling and thinking on by default, text and images in, for a multi-GPU server. | 512 GB | llama.cpp, Ollama, vLLM, SGLang, Transformers | ollama run ornith-1.5:397b |
| MiniMax-M2.5 C Mixture-of-experts model for coding and agent work, trained with reinforcement learning across many environments, tool calling and always-on thinking, text only. | 512 GB | llama.cpp, vLLM, SGLang, LM Studio, Transformers | SAFETENSORS_FAST_GPU=1 vllm serve MiniMaxAI/MiniMax-M2.5 --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser minimax_m2 --reasoning-parser minimax_m2_append_think |
| Ming-3.0-audio-flash E Audio-language model for Chinese and English: speech to text, questions about audio and spoken replies. | 512 GB | hf download inclusionAI/Ming-3.0-audio-flash --local-dir inclusionAI/Ming-3.0-audio-flash && export MODEL_PATH="$PWD/inclusionAI/Ming-3.0-audio-flash" && pip install -r requirements-vllm-omni.txt && MING_OUTPUT_MODE=asr ./scripts/serve_vllm_omni.sh "$MODEL_PATH" |
Needs more than 512 GB
| Model | Needs | Runtimes | Start it |
|---|---|---|---|
| GLM-5.3 B Large model for coding and long-horizon agent work with tool calling and always-on reasoning, text only. | 1024 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve zai-org/GLM-5.3 --kv-cache-dtype fp8 --tensor-parallel-size 8 --speculative-config.method mtp --speculative-config.num_speculative_tokens 5 --tool-call-parser glm47 --reasoning-parser glm47 --enable-auto-tool-choice --served-model-name glm-5.3 |
| Inkling B Large mixture-of-experts generalist for coding, agents, factual recall and audio, with images in and thinking effort set per request. | 1024 GB | vLLM, SGLang, Transformers | sglang serve --trust-remote-code --model-path thinkingmachines/Inkling-NVFP4 --tp 8 --quantization modelopt_fp4 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling |
| Kimi K3 B Large multimodal agent model for coding and long tool-calling sessions, text and images in, with always-on reasoning and JSON Schema output. | 2048 GB | llama.cpp, vLLM, SGLang | sglang serve --trust-remote-code --model-path moonshotai/Kimi-K3 --tp-size 8 --context-length 65536 --mem-fraction-static 0.82 --reasoning-parser kimi_k3 --tool-call-parser kimi_k3 --port 30000 |
| DeepSeek V4.1 Flash C Multimodal mixture-of-experts model for agents and long context, text and images in, with tool calls and a reasoning budget. | 1024 GB | vLLM, SGLang | pip install -r inference/requirements.txt && python convert.py --hf-ckpt-path "$HF_CKPT_PATH" --save-path "$SAVE_PATH" --model-parallel 8 --expert-dtype fp4 --tokenizer-path "$HF_CKPT_PATH" && torchrun --nproc-per-node 8 generate.py --ckpt-path "$SAVE_PATH" --config config.json --interactive |
| DeepSeek-V3.2 C Large text reasoning model trained for tool use and search agents, with thinking and chat modes chosen per request. | 1024 GB | llama.cpp, vLLM, SGLang, Transformers | vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 --tokenizer-mode deepseek_v32 --enable-auto-tool-choice --tool-call-parser deepseek_v32 --reasoning-parser deepseek_v3 |
Accio (Accio-Lab)
Publishes from huggingface.co/Accio-Lab. Checked 2026-10-21.
occamy from 2026-09-03
Accio's agent model line; occamy-1.0 is its first release. Apache-2.0 throughout. The Accio-Lab organisation isn't verified. occamy-1.0's date is the public release in our listing (the repository was created 13 August 2026).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| occamy-1.0 Occamy-1.0 is Accio's open-weight agent model for long multi-step work with tools, files and code, a fine-tune of Qwen3.6-35B-A3B (35B mixture of experts, 3B active) published on Hugging Face under Apache 2.0. (The base's MTP layer is removed in this checkpoint and published separately as occamy-1.0-MTP.) | reasoning (agentic, thinking on by default) | not stated (35,107,181,936 by the Hub's safetensors count) | 2026-09-03 | apache-2.0 | Qwen/Qwen3.6-35B-A3B | Our listing (C) · Download · occamy-1.0-GGUF · occamy-1.0-APEX-GGUF · occamy-1.0-FP8 · occamy-1.0-NVFP4 · occamy-1.0-MLX-3bit · occamy-1.0-MLX-4bit · occamy-1.0-MLX-5bit · occamy-1.0-MLX-6bit · occamy-1.0-MLX-8bit · occamy-1.0-MLX-mxfp4 · occamy-1.0-MLX-mxfp8 · occamy-1.0-MLX-nvfp4 |
| occamy-1.0-MTP | experimental MTP draft head for speculative decoding | not stated (draft head only) | 2026-09-15 | apache-2.0 | Accio-Lab/occamy-1.0 | Download |
Older versions (3 families, 4 models)
Metis from 2026-04-09
Agentic multimodal tool-use research models.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Metis-8B-ColdStart (The Hub counts only 770,288 safetensors parameters for this repository, so its weights may not be complete or may be stored another way; not checked further.) | vision-language (cold-start SFT) | 8B (770,288 by the Hub's safetensors count) | 2026-04-09 | apache-2.0 | Qwen/Qwen3-VL-8B-Instruct | Download |
| Metis-8B-RL | vision-language (RL) | 8B (8,767,123,696 by the Hub's safetensors count) | 2026-04-09 | apache-2.0 | Accio-Lab/Metis-8B-ColdStart (from Qwen/Qwen3-VL-8B-Instruct) | Download |
SwimBird from 2026-02-05
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| SwimBird-8B (No README in the repository (404) and no licence in metadata.) | vision-language (research) | 8B (8,767,123,696 by the Hub's safetensors count) | 2026-02-05 | none in the card's metadata | Qwen/Qwen3-VL-8B-Instruct (not declared; inferred from the SwimBird_Qwen3VL architecture and the same parameter count as Metis-8B-RL) | Download |
yofo from 2026-02-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| yofo-Qwen3-VL-2B-Instruct (Research model for arXiv 2511.16600; no licence in metadata. Named after its base rather than under its own name.) | vision-language (research) | 2B (2,127,532,032 by the Hub's safetensors count) | 2026-02-11 | none in the card's metadata | Qwen/Qwen3-VL-2B-Instruct (not declared; named after it, YOFOQwen3VL architecture) | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: occamy-1.0 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Ai2 (Allen Institute for AI)
Publishes from huggingface.co/allenai. Checked 2026-10-21.
Olmo 3 (3, 3.1) from 2025-09-12
Ai2's fully open (weights, data and code) third generation: 7B and 32B bases, Think and Instruct editions with their SFT and DPO stages, RL-Zero checkpoints trained straight from the 7B base, and the December 2025 3.1 updates (32B Think and Instruct, 7B RL-Zero Math and Code). Apache 2.0 throughout. Ai2 publishes no quantised builds for Olmo 3. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Olmo-3-7B-Think Olmo 3 7B Think is Ai2's fully open 7B reasoning model for maths, code and step-by-step answers, with training data, code and checkpoints published, a 65,536-token context and thinking always on, released in November 2025 under Apache 2.0. | reasoning | 7B (Hub safetensors count 7.30B) | 2025-11-18 | apache-2.0 | Our listing (D) · Download | |
| Olmo-3.1-7B-RL-Zero-Math | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-12-12 | apache-2.0 | Download | |
| Olmo-3.1-7B-RL-Zero-Code | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-12-10 | apache-2.0 | Download | |
| Olmo-3.1-32B-Think | reasoning | 32B (Hub safetensors count 32.23B) | 2025-12-10 | apache-2.0 | Download | |
| Olmo-3.1-32B-Instruct | instruct | 32B (Hub safetensors count 32.23B) | 2025-12-10 | apache-2.0 | Download | |
| Olmo-3-7B-RL-Zero-Mix | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-12-01 | apache-2.0 | Download | |
| Olmo-3-7B-Instruct-DPO | instruct (DPO stage) | 7B (Hub safetensors count 7.30B) | 2025-11-19 | apache-2.0 | Download | |
| Olmo-3-7B-Instruct | instruct | 7B (Hub safetensors count 7.30B) | 2025-11-19 | apache-2.0 | Download | |
| Olmo-3-32B-Think | reasoning | 32B (Hub safetensors count 32.23B) | 2025-11-19 | apache-2.0 | Download | |
| Olmo-3.1-32B-Instruct-SFT | instruct (SFT stage) | 32B (Hub safetensors count 32.23B) | 2025-11-18 | apache-2.0 | Download | |
| Olmo-3.1-32B-Instruct-DPO | instruct (DPO stage) | 32B (Hub safetensors count 32.23B) | 2025-11-18 | apache-2.0 | Download | |
| Olmo-3-7B-Think-DPO | reasoning (DPO stage) | 7B (Hub safetensors count 7.30B) | 2025-11-18 | apache-2.0 | Download | |
| Olmo-3-7B-RL-Zero-Math | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-11-17 | apache-2.0 | Download | |
| Olmo-3-7B-RL-Zero-IF | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-11-17 | apache-2.0 | Download | |
| Olmo-3-7B-RL-Zero-General | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-11-17 | apache-2.0 | Download | |
| Olmo-3-7B-RL-Zero-Code | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.30B) | 2025-11-17 | apache-2.0 | Download | |
| Olmo-3-7B-Instruct-SFT | instruct (SFT stage) | 7B (Hub safetensors count 7.30B) | 2025-11-17 | apache-2.0 | Download | |
| Olmo-3-32B-Think-SFT | reasoning (SFT stage) | 32B (Hub safetensors count 32.23B) | 2025-11-14 | apache-2.0 | Download | |
| Olmo-3-32B-Think-DPO | reasoning (DPO stage) | 32B (Hub safetensors count 32.23B) | 2025-11-14 | apache-2.0 | Download | |
| Olmo-3-1125-32B | base | 32B (Hub safetensors count 32.23B) | 2025-11-04 | apache-2.0 | Download | |
| Olmo-3-7B-Think-SFT | reasoning (SFT stage) | 7B (Hub safetensors count 7.30B) | 2025-10-14 | apache-2.0 | Download | |
| Olmo-3-1025-7B | base | 7B (Hub safetensors count 7.30B) | 2025-09-12 | apache-2.0 | Download |
Older versions (11 families, 95 models)
Olmo Hybrid from 2026-01-28
Ai2's earlier and side lines, newest first. Molmo, Molmo 2 and Tülu are built on other publishers' models (base set from the card), the rest on Ai2's own OLMo weights.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Olmo-Hybrid-Think-SFT-7B | reasoning (SFT stage) | 7B (Hub safetensors count 7.43B) | 2026-02-28 | apache-2.0 | Download | |
| Olmo-Hybrid-Instruct-DPO-7B | instruct (DPO stage) | 7B (Hub safetensors count 7.43B) | 2026-02-20 | apache-2.0 | Download | |
| Olmo-Hybrid-Instruct-SFT-7B | instruct (SFT stage) | 7B (Hub safetensors count 7.43B) | 2026-02-19 | apache-2.0 | Download | |
| Olmo-Hybrid-7B | base | 7B (Hub safetensors count 7.43B) | 2026-01-28 | apache-2.0 | Download |
Molmo 2 (with MolmoPoint and MolmoWeb) from 2025-12-14
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Molmo2-ER | instruct (vision-language) | 4.85B (Hub safetensors count; the name states none) | 2026-05-04 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
| MolmoWeb-8B | instruct (vision-language) | 8B (Hub safetensors count 8.66B) | 2026-03-20 | apache-2.0 | Qwen/Qwen3-8B | Download |
| MolmoWeb-4B | instruct (vision-language) | 4B (Hub safetensors count 4.85B) | 2026-03-20 | apache-2.0 | Qwen/Qwen3-8B | Download |
| MolmoPoint-Vid-4B | instruct (vision-language) | 4B (Hub safetensors count 4.86B) | 2026-03-17 | apache-2.0 | google/siglip-so400m-patch14-384 | Download |
| MolmoPoint-GUI-8B | instruct (vision-language) | 8B (Hub safetensors count 8.68B) | 2026-03-17 | apache-2.0 | Qwen/Qwen3-8B | Download |
| MolmoPoint-8B | instruct (vision-language) | 8B (Hub safetensors count 8.68B) | 2026-03-16 | apache-2.0 | Qwen/Qwen3-8B | Download |
| Molmo2-VideoPoint-4B | instruct (vision-language) | 4B (Hub safetensors count 4.85B) | 2025-12-16 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
| Molmo2-O-7B | instruct (vision-language) | 7B (Hub safetensors count 7.76B) | 2025-12-14 | apache-2.0 | allenai/Olmo-3-7B-Instruct | Download |
| Molmo2-8B | instruct (vision-language) | 8B (Hub safetensors count 8.66B) | 2025-12-14 | apache-2.0 | Qwen/Qwen3-8B | Download |
| Molmo2-4B | instruct (vision-language) | 4B (Hub safetensors count 4.85B) | 2025-12-14 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
Bolmo from 2025-12-10
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Bolmo-7B | base (byte-level) | 7B (Hub safetensors count 7.63B) | 2025-12-13 | apache-2.0 | allenai/Olmo-3-1025-7B | Download |
| Bolmo-1B | base (byte-level) | 1B (Hub safetensors count 1.47B) | 2025-12-10 | apache-2.0 | allenai/OLMo-2-0425-1B | Download |
FlexOlmo from 2025-06-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Flex-public-7B-1T | base (public expert) | 7B (Hub safetensors count 7.30B) | 2025-07-24 | apache-2.0 | Download | |
| FlexOlmo-7x7B-1T | base (mixture of independently trained experts) | 7x7B (Hub safetensors count 33.27B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-reddit-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-pes2o-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-news-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-math-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-creative-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download | |
| Flex-code-2x7B-1T | instruct | 2x7B (Hub safetensors count 11.63B) | 2025-06-11 | apache-2.0 | Download |
Tülu 3 (on Llama 3.1) from 2024-11-18
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama-3.1-Tulu-3.1-8B | instruct | 8B (Hub safetensors count 8.03B) | 2025-02-07 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Tulu-3-405B-DPO | instruct (DPO stage) | 405B | 2025-01-28 | llama3.1 | meta-llama/Llama-3.1-405B | Download |
| Llama-3.1-Tulu-3-405B-SFT | instruct (SFT stage) | 405B | 2025-01-10 | llama3.1 | meta-llama/Llama-3.1-405B | Download |
| Llama-3.1-Tulu-3-405B | instruct | 405B (Hub safetensors count 405.85B) | 2025-01-09 | llama3.1 | meta-llama/Llama-3.1-405B | Download |
| Llama-3.1-Tulu-3-8B-RM | reward model | 8B (Hub safetensors count 7.50B) | 2024-11-20 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Tulu-3-8B-DPO | instruct (DPO stage) | 8B (Hub safetensors count 8.03B) | 2024-11-20 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Tulu-3-8B | instruct | 8B (Hub safetensors count 8.03B) | 2024-11-20 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Tulu-3-70B-DPO | instruct (DPO stage) | 70B (Hub safetensors count 70.55B) | 2024-11-20 | llama3.1 | meta-llama/Llama-3.1-70B | Download |
| Llama-3.1-Tulu-3-70B | instruct | 70B (Hub safetensors count 70.55B) | 2024-11-20 | llama3.1 | meta-llama/Llama-3.1-70B | Download |
| Llama-3.1-Tulu-3-8B-SFT | instruct (SFT stage) | 8B (Hub safetensors count 8.03B) | 2024-11-18 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Tulu-3-70B-SFT | instruct (SFT stage) | 70B (Hub safetensors count 70.55B) | 2024-11-18 | llama3.1 | meta-llama/Llama-3.1-70B | Download |
OLMo 2 from 2024-10-29
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| OLMo-2-0425-1B-RLVR1 | instruct (RLVR stage) | 1B (Hub safetensors count 1.48B) | 2025-04-29 | apache-2.0 | Download | |
| OLMo-2-0425-1B-Instruct | instruct | 1B (Hub safetensors count 1.48B) | 2025-04-29 | apache-2.0 | Download · OLMo-2-0425-1B-Instruct-GGUF | |
| OLMo-2-0425-1B-DPO | instruct (DPO stage) | 1B | 2025-04-28 | apache-2.0 | Download | |
| OLMo-2-0425-1B-SFT | instruct (SFT stage) | 1B | 2025-04-24 | apache-2.0 | Download | |
| OLMo-2-0425-1B | base | 1B (Hub safetensors count 1.48B) | 2025-04-17 | apache-2.0 | Download · OLMo-2-0425-1B-GGUF | |
| OLMo-2-0325-32B-SFT | instruct (SFT stage) | 32B | 2025-03-12 | apache-2.0 | Download | |
| OLMo-2-0325-32B-Instruct | instruct | 32B (Hub safetensors count 32.23B) | 2025-03-12 | apache-2.0 | Download · OLMo-2-0325-32B-Instruct-GGUF | |
| OLMo-2-0325-32B-DPO | instruct (DPO stage) | 32B | 2025-03-12 | apache-2.0 | Download | |
| OLMo-2-0325-32B | base | 32B (Hub safetensors count 32.23B) | 2025-02-23 | apache-2.0 | Download | |
| OLMo-2-1124-7B-SFT | instruct (SFT stage) | 7B | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-7B-RM | reward model | 7B | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-7B-Instruct | instruct | 7B (Hub safetensors count 7.30B) | 2024-12-18 | apache-2.0 | Download · OLMo-2-1124-7B-Instruct-GGUF | |
| OLMo-2-1124-7B-DPO | instruct (DPO stage) | 7B | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B-SFT | instruct (SFT stage) | 13B | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B-RM | reward model | 13B | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B-Instruct-RLVR2 | instruct (RLVR stage) | 13B (Hub safetensors count 13.72B) | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B-Instruct-RLVR1 | instruct (RLVR stage) | 13B (Hub safetensors count 13.72B) | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B-Instruct | instruct | 13B (Hub safetensors count 13.72B) | 2024-12-18 | apache-2.0 | Download · OLMo-2-1124-13B-Instruct-GGUF | |
| OLMo-2-1124-13B-DPO | instruct (DPO stage) | 13B (Hub safetensors count 13.72B) | 2024-12-18 | apache-2.0 | Download | |
| OLMo-2-1124-13B | base | 13B (Hub safetensors count 13.72B) | 2024-11-19 | apache-2.0 | Download · OLMo-2-1124-13B-GGUF | |
| OLMo-2-1124-7B | base | 7B (Hub safetensors count 7.30B) | 2024-10-29 | apache-2.0 | Download · OLMo-2-1124-7B-GGUF |
Molmo from 2024-09-24
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Molmo-7B-O-0924 | instruct (vision-language) | 7B (Hub safetensors count 7.67B) | 2024-09-25 | apache-2.0 | openai/clip-vit-large-patch14-336 | Download |
| Molmo-7B-D-0924 | instruct (vision-language) | 7B (Hub safetensors count 8.02B) | 2024-09-25 | apache-2.0 | Qwen/Qwen2-7B | Download |
| Molmo-72B-0924 | instruct (vision-language) | 72B (Hub safetensors count 73.31B) | 2024-09-25 | apache-2.0 | Qwen/Qwen2-72B | Download |
| MolmoE-1B-0924 | instruct (vision-language) | 7B total, 1B active (MoE), on OLMoE-1B-7B (no safetensors count on the Hub) | 2024-09-24 | apache-2.0 | allenai/OLMoE-1B-7B-0924 | Download |
OLMoE from 2024-07-20
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| OLMoE-1B-7B-0125-SFT | instruct (SFT stage) | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2025-01-27 | apache-2.0 | Download | |
| OLMoE-1B-7B-0125-Instruct | instruct | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2025-01-27 | apache-2.0 | Download · OLMoE-1B-7B-0125-Instruct-GGUF | |
| OLMoE-1B-7B-0125-DPO | instruct (DPO stage) | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2025-01-27 | apache-2.0 | Download | |
| OLMoE-1B-7B-0125 | base | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2025-01-21 | apache-2.0 | Download · OLMoE-1B-7B-0125-GGUF | |
| OLMoE-1B-7B-0924-SFT | instruct (SFT stage) | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2024-08-13 | apache-2.0 | Download | |
| OLMoE-1B-7B-0924-Instruct | instruct | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2024-08-13 | apache-2.0 | Download · OLMoE-1B-7B-0924-Instruct-GGUF | |
| OLMoE-1B-7B-0924 | base | 7B total, 1B active (MoE), per the name 1B-7B (Hub safetensors count 6.92B) | 2024-07-20 | apache-2.0 | Download · OLMoE-1B-7B-0924-GGUF |
OLMo (1) from 2024-01-09
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| OLMo-7B-0724-hf | base | 7B (Hub safetensors count 6.89B) | 2024-07-12 | apache-2.0 | Download | |
| OLMo-7B-0724-Instruct-hf | instruct | 7B (Hub safetensors count 6.89B) | 2024-07-09 | apache-2.0 | Download | |
| OLMo-7B-0724-SFT-hf | instruct (SFT stage) | 7B (Hub safetensors count 6.89B) | 2024-07-08 | apache-2.0 | Download | |
| OLMo-7B-0424-SFT-hf | instruct (SFT stage) | 7B (Hub safetensors count 6.89B) | 2024-07-08 | apache-2.0 | Download | |
| OLMo-7B-0424-Instruct-hf | instruct | 7B (Hub safetensors count 6.89B) | 2024-07-08 | apache-2.0 | Download | |
| OLMo-1B-0724-hf | base | 1B (Hub safetensors count 1.28B) | 2024-06-15 | apache-2.0 | Download | |
| OLMo-7B-SFT | instruct (SFT stage) | 7B | 2024-02-23 | apache-2.0 | Download · OLMo-7B-SFT-hf | |
| OLMo-7B-Instruct | instruct | 7B (Hub safetensors count 6.89B) | 2024-02-23 | apache-2.0 | Download · OLMo-7B-Instruct-hf | |
| OLMo-1B | base | 1B (Hub safetensors count 1.18B) | 2024-01-26 | apache-2.0 | Download · OLMo-1B-hf | |
| OLMo-7B | base | 7B (Hub safetensors count 6.89B) | 2024-01-09 | apache-2.0 | Download · OLMo-7B-hf |
Tülu 2 (on Llama 2, Llama 3 and Llama 3.1) from 2023-11-12
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| llama-3.1-tulu-2-dpo-8b | instruct (DPO stage) | 8B (Hub safetensors count 8.03B) | 2024-08-09 | apache-2.0 | meta-llama/Llama-3.1-8B | Download |
| llama-3.1-tulu-2-dpo-70b | instruct (DPO stage) | 70B (Hub safetensors count 70.55B) | 2024-08-09 | apache-2.0 | meta-llama/Llama-3.1-70B | Download |
| llama-3.1-tulu-2-8b | instruct (SFT) | 8B (Hub safetensors count 8.03B) | 2024-08-09 | apache-2.0 | meta-llama/Llama-3.1-8B | Download |
| llama-3.1-tulu-2-70b | instruct (SFT) | 70B (Hub safetensors count 70.55B) | 2024-08-09 | apache-2.0 | meta-llama/Llama-3.1-70B | Download |
| llama-3-tulu-2-dpo-8b | instruct (DPO stage) | 8B | 2024-06-20 | apache-2.0 | meta-llama/Meta-Llama-3-8B | Download |
| llama-3-tulu-2-dpo-70b | instruct (DPO stage) | 70B (Hub safetensors count 70.55B) | 2024-06-20 | apache-2.0 | meta-llama/Meta-Llama-3-70B | Download |
| llama-3-tulu-2-8b | instruct (SFT) | 8B | 2024-06-20 | apache-2.0 | meta-llama/Meta-Llama-3-8B | Download |
| llama-3-tulu-2-70b | instruct (SFT) | 70B (Hub safetensors count 70.55B) | 2024-06-20 | apache-2.0 | meta-llama/Meta-Llama-3-70B | Download |
| tulu-2-dpo-7b | instruct (DPO stage) | 7B | 2023-11-13 | other (ai2-impact-license-low-risk) | meta-llama/Llama-2-7b-hf | Download |
| tulu-2-dpo-13b | instruct (DPO stage) | 13B (Hub safetensors count 13.02B) | 2023-11-13 | other (ai2-impact-license-low-risk) | meta-llama/Llama-2-13b-hf | Download |
| tulu-2-7b | instruct (SFT) | 7B | 2023-11-13 | none in the card metadata | meta-llama/Llama-2-7b-hf | Download |
| tulu-2-70b | instruct (SFT) | 70B | 2023-11-13 | none in the card metadata | meta-llama/Llama-2-70b-hf | Download |
| tulu-2-13b | instruct (SFT) | 13B | 2023-11-13 | none in the card metadata | meta-llama/Llama-2-13b-hf | Download |
| tulu-2-dpo-70b | instruct (DPO stage) | 70B (Hub safetensors count 68.98B) | 2023-11-12 | other (ai2-impact-license-low-risk) | meta-llama/Llama-2-70b-hf | Download |
Tülu (1, on LLaMA) from 2023-06-07
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Olmo 3 7B Think 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Aleph Alpha
Publishes from huggingface.co/Aleph-Alpha. Checked 2026-10-21.
Kolibri from 2026-10-03
Trained by Aleph Alpha (card describes its own pre-training); no base model named. The card reports a 'Kolibri Base' in its tables but no base repository is published in the organisation (Aleph-Alpha/Kolibri-1-Base answers 401). Release date is from our listing; Hub createdAt is 2026-10-02.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kolibri-1 Kolibri-1 is Aleph Alpha's open-weight mixture-of-experts reasoning model for German and English, 78B parameters with 3.46B active per token, published on Hugging Face in FP8 and BF16 under Apache 2.0 and served with vLLM. | reasoning (instruct) | 78B total, 3.46B active (MoE) | 2026-10-03 | apache-2.0 | Our listing (C) · Download · Kolibri-1 |
Older versions (4 families, 23 models)
Pharia-1 from 2024-08-27
Open Aleph License (Hub id open-aleph-license). Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Pharia-1-LLM-7B-control-hf | instruct (control) | 7B | 2024-08-27 | open-aleph-license | Download · Pharia-1-LLM-7B-control | |
| Pharia-1-LLM-7B-control-aligned-hf | instruct (control, aligned) | 7B | 2024-08-27 | open-aleph-license | Download · Pharia-1-LLM-7B-control-aligned | |
| Pharia-1-Embedding-4608-control-hf | embedding | 7B | 2024-12-17 | open-aleph-license | Aleph-Alpha/Pharia-1-LLM-7B-control (card: 'built on top of Pharia-1-LLM-7B-control') | Download · Pharia-1-Embedding-4608-control |
| Pharia-1-Embedding-4608-control-256 (Separate repository with a -256 suffix; card not read beyond the API fields.) | embedding | 7B | 2024-11-15 | open-aleph-license | Download |
TFree-HAT from 2025-04-25
Tokenizer-free Hierarchical Autoregressive Transformer models. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| tfree-hat-pretrained-7b-base (Lives in the Aleph-Alpha-Research organisation (Aleph-Alpha/tfree-hat-pretrained-7b-base redirects there); licence assumed from its sibling, not read.) | base | 7B | 2025-07-31 | open-aleph-license | Download | |
| llama-tfree-hat-pretrained-7b-dpo | instruct (DPO) | 7B | 2025-07-30 | open-aleph-license | Aleph-Alpha/tfree-hat-pretrained-7b-base | Download |
| llama-3_1-8b-tfree-hat-base | base (Llama 3.1 backbone, HAT tokenizer-free) | 8B | 2025-04-25 | open-aleph-license | meta-llama/Meta-Llama-3.1-8B | Download |
| llama-3_1-8b-tfree-hat-sft | instruct (SFT) | 8B | 2025-04-25 | open-aleph-license | Aleph-Alpha/llama-3_1-8b-tfree-hat-base | Download |
| llama-3_1-8b-tfree-hat-dpo | instruct (DPO) | 8B | 2025-04-25 | open-aleph-license | Aleph-Alpha/llama-3_1-8b-tfree-hat-sft | Download |
| llama-3_1-70b-tfree-hat-sft | instruct (SFT) | 70B | 2025-07-31 | open-aleph-license | Llama 3.1 70B base (card: 'based on the Llama 3.1 70B base model's pre-trained backbone') | Download |
Research checkpoints from 2024-08-28
Not a product family: paper checkpoints (T-Free, u-muP, a standard-parametrisation baseline, MAGMA) in Scaling-format .pt files, grouped here so they aren't lost. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| tfree-research-vocab-32k-fineweb-steps-370k | base (research) | 2024-08-28 | open-aleph-license | Download | ||
| tfree-research-vocab-32k-fineweb-steps-370k-instruct-10k | instruct (research) | 2024-08-28 | open-aleph-license | Download | ||
| umup-research-1b-bf16 | base (research) | 1B | 2024-10-29 | open-aleph-license | Download · umup-research-1b-fp8 | |
| umup-research-3b-bf16 | base (research) | 3B | 2024-10-29 | open-aleph-license | Download · umup-research-3b-fp8 | |
| umup-research-7b-bf16 | base (research) | 7B | 2024-10-28 | open-aleph-license | Download · umup-research-7b-fp8 | |
| sp-baseline-research-1b-bf16 | base (research) | 1B | 2024-10-29 | open-aleph-license | Download | |
| sp-baseline-research-3b-bf16 | base (research) | 3B | 2024-10-29 | open-aleph-license | Download | |
| sp-baseline-research-7b-bf16 | base (research) | 7B | 2024-10-29 | open-aleph-license | Download | |
| magma-research | multimodal (research) | 2025-04-10 | mit | Download |
Aleph-Alpha-GermanWeb classifiers from 2025-04-17
Data-filtering classifiers, not language models. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Aleph-Alpha-GermanWeb-Quality-Classifier-BERT | classifier | 110M | 2025-04-17 | open-aleph-license | google-bert/bert-base-uncased | Download |
| Aleph-Alpha-GermanWeb-Grammar-Classifier-BERT | classifier | 110M | 2025-04-17 | open-aleph-license | google-bert/bert-base-uncased | Download |
| Aleph-Alpha-GermanWeb-Quality-Classifier-fastText | classifier (fastText) | 2025-04-17 | open-aleph-license | Download | ||
| Aleph-Alpha-GermanWeb-Grammar-Classifier-fastText | classifier (fastText) | 2025-04-17 | open-aleph-license | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Kolibri-1 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Alibaba (Qwen)
Publishes from huggingface.co/Qwen. Checked 2026-10-21.
Qwen3 from 2025-04-27
All Qwen3.x point releases (3.5, 3.6, 3.8) and the Qwen3-VL, Qwen3-Coder, Qwen3-Next, Qwen3-Omni, Qwen3-Embedding/Reranker, Qwen3Guard, Qwen3-TTS and Qwen3-ASR lines sit here because the publisher names them Qwen3. WebWorld, Qwen-Drive-1.0-4B and Qwen-AgentWorld-35B-A3B are separately named fine-tunes whose cards name a Qwen3/Qwen3.5 base; they are included with edition "fine-tune (separately named)". The -hf ASR/aligner repos are transformers-format copies and are listed as builds. "released" is the Hub repo creation date (createdAt), which can precede public release by days. Excluded from members: the 14 SAE-Res-* sparse autoencoders (interpretability artefacts, not models) and Qwen-Image-Bench (card names Qwen3.6-27B as base; purpose not checked).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3-0.6B | instruct | 0.6B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-0.6B-Base | Download · Qwen3-0.6B-FP8 · Qwen3-0.6B-GGUF · Qwen3-0.6B-GPTQ-Int8 · Qwen3-0.6B-MLX-4bit · Qwen3-0.6B-MLX-6bit · Qwen3-0.6B-MLX-bf16 · Qwen3-0.6B-MLX-8bit |
| Qwen3-1.7B | instruct | 1.7B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-1.7B-Base | Download · Qwen3-1.7B-FP8 · Qwen3-1.7B-GGUF · Qwen3-1.7B-GPTQ-Int8 · Qwen3-1.7B-MLX-bf16 · Qwen3-1.7B-MLX-6bit · Qwen3-1.7B-MLX-8bit · Qwen3-1.7B-MLX-4bit |
| Qwen3-14B | instruct | 14B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-14B-Base | Download · Qwen3-14B-FP8 · Qwen3-14B-GGUF · Qwen3-14B-AWQ · Qwen3-14B-MLX-8bit · Qwen3-14B-MLX-6bit · Qwen3-14B-MLX-4bit · Qwen3-14B-MLX-bf16 |
| Qwen3-235B-A22B | instruct | 235B-A22B | 2025-04-27 | apache-2.0 | Download · Qwen3-235B-A22B-FP8 · Qwen3-235B-A22B-GPTQ-Int4 · Qwen3-235B-A22B-GGUF · Qwen3-235B-A22B-MLX-bf16 · Qwen3-235B-A22B-MLX-4bit · Qwen3-235B-A22B-MLX-6bit · Qwen3-235B-A22B-MLX-8bit | |
| Qwen3-30B-A3B | instruct | 30B-A3B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-30B-A3B-Base | Download · Qwen3-30B-A3B-FP8 · Qwen3-30B-A3B-GGUF · Qwen3-30B-A3B-GPTQ-Int4 · Qwen3-30B-A3B-MLX-bf16 · Qwen3-30B-A3B-MLX-8bit · Qwen3-30B-A3B-MLX-6bit · Qwen3-30B-A3B-MLX-4bit |
| Qwen3-32B | instruct | 32B | 2025-04-27 | apache-2.0 | Download · Qwen3-32B-FP8 · Qwen3-32B-GGUF · Qwen3-32B-AWQ · Qwen3-32B-MLX-bf16 · Qwen3-32B-MLX-8bit · Qwen3-32B-MLX-6bit · Qwen3-32B-MLX-4bit | |
| Qwen3-4B | instruct | 4B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-4B-Base | Download · Qwen3-4B-FP8 · Qwen3-4B-GGUF · Qwen3-4B-AWQ · Qwen3-4B-MLX-bf16 · Qwen3-4B-MLX-8bit · Qwen3-4B-MLX-6bit · Qwen3-4B-MLX-4bit |
| Qwen3-8B | instruct | 8B | 2025-04-27 | apache-2.0 | Qwen/Qwen3-8B-Base | Download · Qwen3-8B-FP8 · Qwen3-8B-AWQ · Qwen3-8B-GGUF · Qwen3-8B-MLX-6bit · Qwen3-8B-MLX-4bit · Qwen3-8B-MLX-8bit · Qwen3-8B-MLX-bf16 |
| Qwen3-0.6B-Base | base | 0.6B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-1.7B-Base | base | 1.7B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-14B-Base | base | 14B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-30B-A3B-Base | base | 30B-A3B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-4B-Base | base | 4B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-8B-Base | base | 8B | 2025-04-28 | apache-2.0 | Download | |
| Qwen3-Reranker-0.6B | reranker | 0.6B | 2025-05-29 | apache-2.0 | Qwen/Qwen3-0.6B-Base | Download |
| Qwen3-Reranker-8B | reranker | 8B | 2025-05-29 | apache-2.0 | Qwen/Qwen3-8B-Base | Download |
| Qwen3-Embedding-0.6B Qwen3-Embedding-0.6B is Alibaba's open-weight text embedding model of 0.6B parameters, built on Qwen3-0.6B-Base, for retrieval, classification and clustering in over 100 languages and code, with 1,024-dimension vectors and a 32K context, under Apache 2.0. | embedding | 0.6B | 2025-06-03 | apache-2.0 | Qwen/Qwen3-0.6B-Base | Our listing (B) · Download · Qwen3-Embedding-0.6B-GGUF |
| Qwen3-Embedding-4B Qwen3-Embedding-4B is Alibaba's open-weight text embedding model built on Qwen3-4B-Base, returning vectors of up to 2,560 dimensions for retrieval, code search and clustering in over 100 languages, under Apache 2.0. | embedding | 4B | 2025-06-03 | apache-2.0 | Qwen/Qwen3-4B-Base | Our listing (BB) · Download · Qwen3-Embedding-4B-GGUF |
| Qwen3-Embedding-8B Qwen3-Embedding-8B is Alibaba's open-weight text embedding model built on Qwen3-8B-Base, returning vectors of up to 4,096 dimensions for retrieval, code search and clustering in over 100 languages, under Apache 2.0. | embedding | 8B | 2025-06-03 | apache-2.0 | Qwen/Qwen3-8B-Base | Our listing (BB) · Download · Qwen3-Embedding-8B-GGUF |
| Qwen3-Reranker-4B | reranker | 4B | 2025-06-03 | apache-2.0 | Qwen/Qwen3-4B-Base | Download |
| Qwen3-235B-A22B-Instruct-2507 | instruct | 235B-A22B | 2025-07-21 | apache-2.0 | Download · Qwen3-235B-A22B-Instruct-2507-FP8 | |
| Qwen3-Coder-480B-A35B-Instruct | instruct | 480B-A35B | 2025-07-22 | apache-2.0 | Download · Qwen3-Coder-480B-A35B-Instruct-FP8 | |
| Qwen3-235B-A22B-Thinking-2507 | reasoning | 235B-A22B | 2025-07-25 | apache-2.0 | Download · Qwen3-235B-A22B-Thinking-2507-FP8 | |
| Qwen3-30B-A3B-Instruct-2507 Qwen3-30B-A3B-Instruct-2507 is Alibaba's open-weight mixture-of-experts text model, 30.5B parameters with 3.3B active, the non-thinking edition of the July 2025 Qwen3 update, with a 262,144-token context, published in BF16 and FP8 under Apache 2.0. | instruct | 30B-A3B | 2025-07-28 | apache-2.0 | Our listing (C) · Download · Qwen3-30B-A3B-Instruct-2507-FP8 | |
| Qwen3-30B-A3B-Thinking-2507 | reasoning | 30B-A3B | 2025-07-29 | apache-2.0 | Download · Qwen3-30B-A3B-Thinking-2507-FP8 | |
| Qwen3-Coder-30B-A3B-Instruct Qwen3-Coder-30B-A3B-Instruct is Alibaba's open-weight mixture-of-experts coding model, 30.5B parameters with 3.3B active, for agentic coding and tool use, with a 262,144-token context, published in BF16 and FP8 under Apache 2.0. | instruct | 30B-A3B | 2025-07-31 | apache-2.0 | Our listing (C) · Download · Qwen3-Coder-30B-A3B-Instruct-FP8 | |
| Qwen3-4B-Instruct-2507 Qwen3-4B-Instruct-2507 is Alibaba's open-weight dense text model of 4B parameters, the non-thinking edition of the July 2025 Qwen3 update, with a 262,144-token context, published on Hugging Face in BF16 and FP8 under Apache 2.0. | instruct | 4B | 2025-08-05 | apache-2.0 | Our listing (C) · Download · Qwen3-4B-Instruct-2507-FP8 | |
| Qwen3-4B-Thinking-2507 | reasoning | 4B | 2025-08-05 | apache-2.0 | Download · Qwen3-4B-Thinking-2507-FP8 | |
| Qwen3-Next-80B-A3B-Instruct | instruct | 80B-A3B | 2025-09-09 | apache-2.0 | Download · Qwen3-Next-80B-A3B-Instruct-FP8 · Qwen3-Next-80B-A3B-Instruct-GGUF | |
| Qwen3-Next-80B-A3B-Thinking | reasoning | 80B-A3B | 2025-09-09 | apache-2.0 | Download · Qwen3-Next-80B-A3B-Thinking-FP8 · Qwen3-Next-80B-A3B-Thinking-GGUF | |
| Qwen3-Omni-30B-A3B-Captioner | captioner | 30B-A3B | 2025-09-15 | apache-2.0 | Download | |
| Qwen3-Omni-30B-A3B-Thinking | reasoning | 30B-A3B | 2025-09-15 | apache-2.0 | Download | |
| Qwen3-Omni-30B-A3B-Instruct Qwen3-Omni-30B-A3B-Instruct is Alibaba's open-weight omni-modal mixture of experts, a Thinker that takes text, images, audio and video and a Talker that answers in text or speech, about 35B parameters with 3B active, published in BF16 under Apache 2.0. | instruct | 30B-A3B | 2025-09-20 | apache-2.0 | Our listing (E) · Download | |
| Qwen3-VL-235B-A22B-Instruct | instruct | 235B-A22B | 2025-09-22 | apache-2.0 | Download · Qwen3-VL-235B-A22B-Instruct-FP8 · Qwen3-VL-235B-A22B-Instruct-GGUF | |
| Qwen3-VL-235B-A22B-Thinking | reasoning | 235B-A22B | 2025-09-22 | apache-2.0 | Download · Qwen3-VL-235B-A22B-Thinking-FP8 · Qwen3-VL-235B-A22B-Thinking-GGUF | |
| Qwen3Guard-Gen-0.6B | guard (safety classifier) | 0.6B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| Qwen3Guard-Gen-4B | guard (safety classifier) | 4B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-4B | Download |
| Qwen3Guard-Gen-8B | guard (safety classifier) | 8B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-8B | Download |
| Qwen3Guard-Stream-0.6B | guard (safety classifier) | 0.6B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| Qwen3Guard-Stream-4B | guard (safety classifier) | 4B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-4B | Download |
| Qwen3Guard-Stream-8B | guard (safety classifier) | 8B | 2025-09-23 | apache-2.0 | Qwen/Qwen3-8B | Download |
| Qwen3-4B-SafeRL | instruct | 4B | 2025-09-30 | apache-2.0 | Qwen/Qwen3-4B | Download |
| Qwen3-VL-30B-A3B-Instruct Qwen3-VL-30B-A3B-Instruct is Alibaba's open-weight vision-language mixture of experts with about 31B parameters and 3B active, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | 30B-A3B | 2025-09-30 | apache-2.0 | Our listing (C) · Download · Qwen3-VL-30B-A3B-Instruct-FP8 · Qwen3-VL-30B-A3B-Instruct-GGUF | |
| Qwen3-VL-30B-A3B-Thinking | reasoning | 30B-A3B | 2025-09-30 | apache-2.0 | Download · Qwen3-VL-30B-A3B-Thinking-FP8 · Qwen3-VL-30B-A3B-Thinking-GGUF | |
| Qwen3-VL-4B-Instruct Qwen3-VL-4B-Instruct is Alibaba's open-weight dense vision-language model of about 4B parameters, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | 4B | 2025-10-11 | apache-2.0 | Our listing (D) · Download · Qwen3-VL-4B-Instruct-FP8 · Qwen3-VL-4B-Instruct-GGUF | |
| Qwen3-VL-4B-Thinking | reasoning | 4B | 2025-10-11 | apache-2.0 | Download · Qwen3-VL-4B-Thinking-FP8 · Qwen3-VL-4B-Thinking-GGUF | |
| Qwen3-VL-8B-Instruct Qwen3-VL-8B-Instruct is Alibaba's open-weight dense vision-language model of about 8B parameters, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | 8B | 2025-10-11 | apache-2.0 | Our listing (C) · Download · Qwen3-VL-8B-Instruct-FP8 · Qwen3-VL-8B-Instruct-GGUF | |
| Qwen3-VL-8B-Thinking | reasoning | 8B | 2025-10-11 | apache-2.0 | Download · Qwen3-VL-8B-Thinking-FP8 · Qwen3-VL-8B-Thinking-GGUF | |
| Qwen3-VL-2B-Instruct Qwen3-VL-2B-Instruct is Alibaba's open-weight dense vision-language model of about 2B parameters, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | 2B | 2025-10-19 | apache-2.0 | Our listing (C) · Download · Qwen3-VL-2B-Instruct-FP8 · Qwen3-VL-2B-Instruct-GGUF | |
| Qwen3-VL-2B-Thinking | reasoning | 2B | 2025-10-19 | apache-2.0 | Download · Qwen3-VL-2B-Thinking-FP8 · Qwen3-VL-2B-Thinking-GGUF | |
| Qwen3-VL-32B-Instruct Qwen3-VL-32B-Instruct is Alibaba's open-weight dense vision-language model of about 32B parameters, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | 32B | 2025-10-19 | apache-2.0 | Our listing (C) · Download · Qwen3-VL-32B-Instruct-FP8 · Qwen3-VL-32B-Instruct-GGUF | |
| Qwen3-VL-32B-Thinking | reasoning | 32B | 2025-10-19 | apache-2.0 | Download · Qwen3-VL-32B-Thinking-FP8 · Qwen3-VL-32B-Thinking-GGUF | |
| Qwen3-VL-Embedding-2B Qwen3-VL-Embedding-2B is Alibaba's open-weight multimodal embedding model of 2.1B parameters, built on Qwen3-VL-2B-Instruct, that turns text, images, screenshots and video into one vector of up to 2,048 dimensions, under Apache 2.0. | embedding | 2B | 2026-01-07 | apache-2.0 | Qwen/Qwen3-VL-2B-Instruct | Our listing (C) · Download |
| Qwen3-VL-Embedding-8B Qwen3-VL-Embedding-8B is Alibaba's open-weight multimodal embedding model built on Qwen3-VL-8B-Instruct, mapping text, images, screenshots, video or a mix of them to one vector of up to 4,096 dimensions, under Apache 2.0. | embedding | 8B | 2026-01-07 | apache-2.0 | Qwen/Qwen3-VL-8B-Instruct | Our listing (B) · Download |
| Qwen3-VL-Reranker-2B | reranker | 2B | 2026-01-07 | apache-2.0 | Qwen/Qwen3-VL-2B-Instruct | Download |
| Qwen3-VL-Reranker-8B | reranker | 8B | 2026-01-07 | apache-2.0 | Qwen/Qwen3-VL-8B-Instruct | Download |
| Qwen3-TTS-12Hz-0.6B-Base | text-to-speech | 0.6B | 2026-01-21 | apache-2.0 | Download | |
| Qwen3-TTS-12Hz-0.6B-CustomVoice | text-to-speech | 0.6B | 2026-01-21 | apache-2.0 | Download | |
| Qwen3-TTS-12Hz-1.7B-Base | text-to-speech | 1.7B | 2026-01-21 | apache-2.0 | Download | |
| Qwen3-TTS-12Hz-1.7B-CustomVoice Qwen3-TTS 12Hz 1.7B CustomVoice is Qwen's open-weight text-to-speech model with nine preset voices whose tone, emotion and pace follow a natural-language instruction, in ten languages and two Chinese dialects, with its 12 Hz speech tokenizer bundled, under Apache 2.0. | text-to-speech | 1.7B | 2026-01-21 | apache-2.0 | Our listing (C) · Download | |
| Qwen3-TTS-12Hz-1.7B-VoiceDesign | text-to-speech | 1.7B | 2026-01-21 | apache-2.0 | Download | |
| Qwen3-TTS-Tokenizer-12Hz | speech tokenizer | not in name | 2026-01-21 | apache-2.0 | Download | |
| Qwen3-ASR-0.6B Qwen3-ASR-0.6B is Alibaba's compact open-weight speech recognition model for 30 languages and 22 Chinese dialects, with language identification, context hints and streaming, published on Hugging Face in BF16 under Apache 2.0. | speech recognition | 0.6B | 2026-01-28 | apache-2.0 | Our listing (C) · Download · Qwen3-ASR-0.6B-hf | |
| Qwen3-ASR-1.7B Qwen3-ASR-1.7B is Alibaba's open-weight speech recognition model for 30 languages and 22 Chinese dialects, with language identification, context hints and streaming, published on Hugging Face in BF16 under Apache 2.0. | speech recognition | 1.7B | 2026-01-28 | apache-2.0 | Our listing (B) · Download · Qwen3-ASR-1.7B-hf | |
| Qwen3-ForcedAligner-0.6B | forced aligner | 0.6B | 2026-01-28 | apache-2.0 | Download · Qwen3-ForcedAligner-0.6B-hf | |
| Qwen3-Coder-Next Qwen3-Coder-Next is Alibaba's open-weight coding-agent model, an 80B mixture-of-experts with 3B active and no thinking mode, trained for tool calling across many agent scaffolds with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0. | instruct | not in name; Hub safetensors total 79.7B | 2026-01-30 | apache-2.0 | Our listing (C) · Download · Qwen3-Coder-Next-FP8 · Qwen3-Coder-Next-GGUF | |
| Qwen3-Coder-Next-Base | base | not in name; Hub safetensors total 79.7B | 2026-02-01 | apache-2.0 | Download | |
| WebWorld-14B | fine-tune (separately named) | 14B | 2026-02-13 | apache-2.0 | Qwen/Qwen3-8B | Download |
| WebWorld-32B | fine-tune (separately named) | 32B | 2026-02-13 | apache-2.0 | Qwen/Qwen3-8B | Download |
| WebWorld-8B | fine-tune (separately named) | 8B | 2026-02-13 | apache-2.0 | Qwen/Qwen3-8B | Download |
| Qwen3.5-397B-A17B Qwen3.5-397B-A17B is Alibaba's largest open-weight Qwen3.5 model, a mixture-of-experts vision-language model of 397B parameters with 17B active, thinking on by default and a 262,144-token context, published on Hugging Face in BF16, FP8 and GPTQ 4-bit under Apache 2.0. | instruct | 397B-A17B | 2026-02-16 | apache-2.0 | Our listing (B) · Download · Qwen3.5-397B-A17B-FP8 · Qwen3.5-397B-A17B-GPTQ-Int4 | |
| Qwen3.5-122B-A10B Qwen3.5-122B-A10B is Alibaba's mid-size open-weight mixture-of-experts vision-language model, 122B parameters with 10B active, thinking on by default and a 262,144-token context, published on Hugging Face in BF16, FP8 and GPTQ 4-bit under Apache 2.0. | instruct | 122B-A10B | 2026-02-24 | apache-2.0 | Our listing (B) · Download · Qwen3.5-122B-A10B-FP8 · Qwen3.5-122B-A10B-GPTQ-Int4 | |
| Qwen3.5-27B | instruct | 27B | 2026-02-24 | apache-2.0 | Download · Qwen3.5-27B-FP8 · Qwen3.5-27B-GPTQ-Int4 | |
| Qwen3.5-35B-A3B | instruct | 35B-A3B | 2026-02-24 | apache-2.0 | Qwen/Qwen3.5-35B-A3B-Base | Download · Qwen3.5-35B-A3B-FP8 · Qwen3.5-35B-A3B-GPTQ-Int4 |
| Qwen3.5-35B-A3B-Base | base | 35B-A3B | 2026-02-24 | apache-2.0 | Download | |
| Qwen3.5-9B-Base | base | 9B | 2026-02-26 | apache-2.0 | Download | |
| Qwen3.5-4B Qwen3.5-4B is Alibaba's small open-weight vision-language model, 4B language parameters with a vision encoder, thinking on by default and a 262,144-token context, published on Hugging Face in BF16 under Apache 2.0. | instruct | 4B | 2026-02-27 | apache-2.0 | Qwen/Qwen3.5-4B-Base | Our listing (C) · Download |
| Qwen3.5-4B-Base | base | 4B | 2026-02-27 | apache-2.0 | Download | |
| Qwen3.5-9B Qwen3.5-9B is Alibaba's open-weight dense vision-language model of 9B parameters, with thinking on by default and a 262,144-token context, published on Hugging Face in BF16 under Apache 2.0. | instruct | 9B | 2026-02-27 | apache-2.0 | Qwen/Qwen3.5-9B-Base | Our listing (C) · Download |
| Qwen3.5-0.8B Qwen3.5-0.8B is the smallest model in Alibaba's Qwen3.5 family, 0.8B language parameters with a vision encoder, thinking off by default and a 262,144-token context, published on Hugging Face in BF16 under Apache 2.0 for prototyping, fine-tuning and research. | instruct | 0.8B | 2026-02-28 | apache-2.0 | Qwen/Qwen3.5-0.8B-Base | Our listing (D) · Download |
| Qwen3.5-0.8B-Base | base | 0.8B | 2026-02-28 | apache-2.0 | Download | |
| Qwen3.5-2B Qwen3.5-2B is a small model in Alibaba's Qwen3.5 family, 2B language parameters with a vision encoder, thinking off by default and a 262,144-token context, published on Hugging Face in BF16 under Apache 2.0 for prototyping, fine-tuning and research. | instruct | 2B | 2026-02-28 | apache-2.0 | Qwen/Qwen3.5-2B-Base | Our listing (D) · Download |
| Qwen3.5-2B-Base | base | 2B | 2026-02-28 | apache-2.0 | Download | |
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B is Alibaba's open-weight mixture-of-experts vision-language model, 35B parameters with 3B active and a 262,144-token context, published on Hugging Face in BF16 and FP8 under Apache 2.0. | instruct | 35B-A3B | 2026-04-15 | apache-2.0 | Our listing (B) · Download · Qwen3.6-35B-A3B-FP8 | |
| Qwen3.6-27B Qwen3.6-27B is Alibaba's open-weight dense vision-language model of 27B parameters, tuned for agentic coding, with thinking on by default, preserved thinking across turns and a 262,144-token context, published on Hugging Face in BF16 and FP8 under Apache 2.0. | instruct | 27B | 2026-04-21 | apache-2.0 | Our listing (B) · Download · Qwen3.6-27B-FP8 | |
| Qwen-AgentWorld-35B-A3B | fine-tune (separately named) | 35B-A3B | 2026-06-22 | apache-2.0 | Qwen/Qwen3.5-35B-A3B-Base | Download |
| Qwen3.8-27B Qwen3.8-27B is Alibaba's open-weight dense vision-language model of 27B parameters, with thinking on by default and a 262,144-token context, published on Hugging Face in BF16 and FP8 under Apache 2.0. | instruct | 27B | 2026-08-05 | apache-2.0 | Our listing (B) · Download · Qwen3.8-27B-FP8 | |
| Qwen3.8-2.4T-A95B | instruct | 2.4T-A95B | 2026-08-08 | other: qwen3.8-max | Download · Qwen3.8-2.4T-A95B-FP8 | |
| Qwen3.8-Flash-Next Qwen3.8-Flash-Next is Alibaba's open-weight multimodal mixture-of-experts model and an experimental preview of the Qwen4 architecture, 125B parameters with 6B active plus a 51B n-gram embedding table, published on Hugging Face in BF16 and FP8. | instruct | 125B total, 6B active, plus a 51B n-gram embedding table (publisher card); Hub safetensors total 180.0B | 2026-08-24 | other: qwen-community-1.0 | Our listing (C) · Download · Qwen3.8-Flash-Next-FP8 | |
| Qwen-Drive-1.0-4B | fine-tune (separately named) | 4B | 2026-08-27 | apache-2.0 | Qwen/Qwen3.5-4B | Download |
Older versions (4 families, 109 models)
Qwen2.5 from 2024-09-15
Includes QwQ-32B and QwQ-32B-Preview (cards name Qwen2.5-32B / Qwen2.5-32B-Instruct as base) and WorldPM-72B reward models (card names Qwen2.5-72B), which the publisher brands separately. "released" is the Hub repo creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen2.5-0.5B | base | 0.5B | 2024-09-15 | apache-2.0 | Download | |
| Qwen2.5-1.5B | base | 1.5B | 2024-09-15 | apache-2.0 | Download | |
| Qwen2.5-14B | base | 14B | 2024-09-15 | apache-2.0 | Download | |
| Qwen2.5-32B | base | 32B | 2024-09-15 | apache-2.0 | Download | |
| Qwen2.5-3B | base | 3B | 2024-09-15 | other: qwen-research | Download | |
| Qwen2.5-72B | base | 72B | 2024-09-15 | other: qwen | Download | |
| Qwen2.5-7B | base | 7B | 2024-09-15 | apache-2.0 | Download | |
| Qwen2.5-0.5B-Instruct | instruct | 0.5B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-0.5B | Download · Qwen2.5-0.5B-Instruct-GPTQ-Int4 · Qwen2.5-0.5B-Instruct-GPTQ-Int8 · Qwen2.5-0.5B-Instruct-AWQ · Qwen2.5-0.5B-Instruct-GGUF |
| Qwen2.5-14B-Instruct | instruct | 14B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-14B | Download · Qwen2.5-14B-Instruct-GPTQ-Int4 · Qwen2.5-14B-Instruct-GPTQ-Int8 · Qwen2.5-14B-Instruct-AWQ · Qwen2.5-14B-Instruct-GGUF |
| Qwen2.5-72B-Instruct | instruct | 72B | 2024-09-16 | other: qwen | Qwen/Qwen2.5-72B | Download · Qwen2.5-72B-Instruct-GPTQ-Int4 · Qwen2.5-72B-Instruct-GPTQ-Int8 · Qwen2.5-72B-Instruct-AWQ · Qwen2.5-72B-Instruct-GGUF |
| Qwen2.5-7B-Instruct | instruct | 7B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-7B | Download · Qwen2.5-7B-Instruct-GPTQ-Int4 · Qwen2.5-7B-Instruct-GPTQ-Int8 · Qwen2.5-7B-Instruct-AWQ · Qwen2.5-7B-Instruct-GGUF |
| Qwen2.5-Coder-7B | base | 7B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-7B | Download |
| Qwen2.5-Math-1.5B | base | 1.5B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download |
| Qwen2.5-Math-1.5B-Instruct | instruct | 1.5B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-Math-1.5B | Download |
| Qwen2.5-Math-72B | base | 72B | 2024-09-16 | other: qwen | Qwen/Qwen2.5-72B | Download |
| Qwen2.5-Math-72B-Instruct | instruct | 72B | 2024-09-16 | other: qwen | Qwen/Qwen2.5-Math-72B | Download |
| Qwen2.5-Math-7B | base | 7B | 2024-09-16 | apache-2.0 | Qwen/Qwen2.5-7B | Download |
| Qwen2.5-1.5B-Instruct | instruct | 1.5B | 2024-09-17 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download · Qwen2.5-1.5B-Instruct-GPTQ-Int4 · Qwen2.5-1.5B-Instruct-GPTQ-Int8 · Qwen2.5-1.5B-Instruct-AWQ · Qwen2.5-1.5B-Instruct-GGUF |
| Qwen2.5-32B-Instruct | instruct | 32B | 2024-09-17 | apache-2.0 | Qwen/Qwen2.5-32B | Download · Qwen2.5-32B-Instruct-GPTQ-Int4 · Qwen2.5-32B-Instruct-GPTQ-Int8 · Qwen2.5-32B-Instruct-AWQ · Qwen2.5-32B-Instruct-GGUF |
| Qwen2.5-3B-Instruct | instruct | 3B | 2024-09-17 | other: qwen-research | Qwen/Qwen2.5-3B | Download · Qwen2.5-3B-Instruct-GPTQ-Int4 · Qwen2.5-3B-Instruct-GPTQ-Int8 · Qwen2.5-3B-Instruct-AWQ · Qwen2.5-3B-Instruct-GGUF |
| Qwen2.5-Coder-7B-Instruct | instruct | 7B | 2024-09-17 | apache-2.0 | Qwen/Qwen2.5-Coder-7B | Download · Qwen2.5-Coder-7B-Instruct-GGUF · Qwen2.5-Coder-7B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-7B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-7B-Instruct-AWQ |
| Qwen2.5-Math-RM-72B | reward model | 72B | 2024-09-17 | other: qwen | Qwen/Qwen2.5-Math-72B-Instruct | Download |
| Qwen2.5-Coder-1.5B | base | 1.5B | 2024-09-18 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download |
| Qwen2.5-Coder-1.5B-Instruct | instruct | 1.5B | 2024-09-18 | apache-2.0 | Qwen/Qwen2.5-Coder-1.5B | Download · Qwen2.5-Coder-1.5B-Instruct-GGUF · Qwen2.5-Coder-1.5B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-1.5B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-1.5B-Instruct-AWQ |
| Qwen2.5-Math-7B-Instruct | instruct | 7B | 2024-09-19 | apache-2.0 | Qwen/Qwen2.5-Math-7B | Download |
| Qwen2.5-Coder-0.5B-Instruct | instruct | 0.5B | 2024-11-06 | apache-2.0 | Qwen/Qwen2.5-Coder-0.5B | Download · Qwen2.5-Coder-0.5B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-0.5B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-0.5B-Instruct-AWQ · Qwen2.5-Coder-0.5B-Instruct-GGUF |
| Qwen2.5-Coder-14B-Instruct | instruct | 14B | 2024-11-06 | apache-2.0 | Qwen/Qwen2.5-Coder-14B | Download · Qwen2.5-Coder-14B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-14B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-14B-Instruct-AWQ · Qwen2.5-Coder-14B-Instruct-GGUF |
| Qwen2.5-Coder-32B-Instruct | instruct | 32B | 2024-11-06 | apache-2.0 | Qwen/Qwen2.5-Coder-32B | Download · Qwen2.5-Coder-32B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-32B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-32B-Instruct-AWQ · Qwen2.5-Coder-32B-Instruct-GGUF |
| Qwen2.5-Coder-3B-Instruct | instruct | 3B | 2024-11-06 | other: qwen-research | Qwen/Qwen2.5-Coder-3B | Download · Qwen2.5-Coder-3B-Instruct-GPTQ-Int8 · Qwen2.5-Coder-3B-Instruct-GPTQ-Int4 · Qwen2.5-Coder-3B-Instruct-AWQ · Qwen2.5-Coder-3B-Instruct-GGUF |
| Qwen2.5-Coder-0.5B | base | 0.5B | 2024-11-08 | apache-2.0 | Qwen/Qwen2.5-0.5B | Download |
| Qwen2.5-Coder-14B | base | 14B | 2024-11-08 | apache-2.0 | Qwen/Qwen2.5-14B | Download |
| Qwen2.5-Coder-32B | base | 32B | 2024-11-08 | apache-2.0 | Qwen/Qwen2.5-32B | Download |
| Qwen2.5-Coder-3B | base | 3B | 2024-11-08 | other: qwen-research | Qwen/Qwen2.5-3B | Download |
| QwQ-32B-Preview | reasoning | 32B | 2024-11-27 | apache-2.0 | Qwen/Qwen2.5-32B-Instruct | Download |
| Qwen2.5-Math-7B-PRM800K | reward model | 7B | 2025-01-13 | other: qwen | Qwen/Qwen2.5-Math-7B-Instruct | Download |
| Qwen2.5-Math-PRM-72B | reward model | 72B | 2025-01-13 | other: qwen | Qwen/Qwen2.5-Math-72B-Instruct | Download |
| Qwen2.5-Math-PRM-7B | reward model | 7B | 2025-01-13 | other: qwen | Qwen/Qwen2.5-Math-7B-Instruct | Download |
| Qwen2.5-14B-Instruct-1M | instruct | 14B | 2025-01-23 | apache-2.0 | Qwen/Qwen2.5-14B | Download |
| Qwen2.5-7B-Instruct-1M | instruct | 7B | 2025-01-23 | apache-2.0 | Qwen/Qwen2.5-7B | Download |
| Qwen2.5-VL-3B-Instruct | instruct | 3B | 2025-01-26 | other: qwen-research | Download · Qwen2.5-VL-3B-Instruct-AWQ | |
| Qwen2.5-VL-7B-Instruct | instruct | 7B | 2025-01-26 | apache-2.0 | Download · Qwen2.5-VL-7B-Instruct-AWQ | |
| Qwen2.5-VL-72B-Instruct | instruct | 72B | 2025-01-27 | other: qwen | Download · Qwen2.5-VL-72B-Instruct-AWQ | |
| QwQ-32B | reasoning | 32B | 2025-03-05 | apache-2.0 | Qwen/Qwen2.5-32B | Download · QwQ-32B-AWQ · QwQ-32B-GGUF |
| Qwen2.5-VL-32B-Instruct | instruct | 32B | 2025-03-21 | apache-2.0 | Download · Qwen2.5-VL-32B-Instruct-AWQ | |
| Qwen2.5-Omni-7B | instruct | 7B | 2025-03-22 | apache-2.0 | Download · Qwen2.5-Omni-7B-AWQ · Qwen2.5-Omni-7B-GPTQ-Int4 | |
| Qwen2.5-Omni-3B | instruct | 3B | 2025-04-30 | other: qwen-research | Download | |
| WorldPM-72B | reward model | 72B | 2025-05-16 | apache-2.0 | Qwen/Qwen2.5-72B | Download |
| WorldPM-72B-HelpSteer2 | reward model | 72B | 2025-05-16 | apache-2.0 | Qwen/WorldPM-72B | Download |
| WorldPM-72B-RLHFLow | reward model | 72B | 2025-05-16 | apache-2.0 | Qwen/WorldPM-72B | Download |
| WorldPM-72B-UltraFeedback | reward model | 72B | 2025-05-16 | apache-2.0 | Qwen/WorldPM-72B | Download |
Qwen2 from 2024-05-22
Includes QVQ-72B-Preview (card names Qwen2-VL-72B as base). "released" is the Hub repo creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen2-57B-A14B | base | 57B-A14B | 2024-05-22 | apache-2.0 | Download | |
| Qwen2-72B | base | 72B | 2024-05-22 | other: tongyi-qianwen | Download | |
| Qwen2-72B-Instruct | instruct | 72B | 2024-05-28 | other: tongyi-qianwen | Qwen/Qwen2-72B | Download · Qwen2-72B-Instruct-GPTQ-Int4 · Qwen2-72B-Instruct-GPTQ-Int8 · Qwen2-72B-Instruct-AWQ · Qwen2-72B-Instruct-GGUF |
| Qwen2-0.5B | base | 0.5B | 2024-05-31 | apache-2.0 | Download | |
| Qwen2-1.5B | base | 1.5B | 2024-05-31 | apache-2.0 | Download | |
| Qwen2-0.5B-Instruct | instruct | 0.5B | 2024-06-03 | apache-2.0 | Qwen/Qwen2-0.5B | Download · Qwen2-0.5B-Instruct-GPTQ-Int4 · Qwen2-0.5B-Instruct-GPTQ-Int8 · Qwen2-0.5B-Instruct-AWQ · Qwen2-0.5B-Instruct-MLX · Qwen2-0.5B-Instruct-GGUF |
| Qwen2-1.5B-Instruct | instruct | 1.5B | 2024-06-03 | apache-2.0 | Download · Qwen2-1.5B-Instruct-GPTQ-Int4 · Qwen2-1.5B-Instruct-GPTQ-Int8 · Qwen2-1.5B-Instruct-AWQ · Qwen2-1.5B-Instruct-MLX · Qwen2-1.5B-Instruct-GGUF | |
| Qwen2-57B-A14B-Instruct | instruct | 57B-A14B | 2024-06-04 | apache-2.0 | Qwen/Qwen2-57B-A14B | Download · Qwen2-57B-A14B-Instruct-GPTQ-Int4 · Qwen2-57B-A14B-Instruct-GGUF |
| Qwen2-7B | base | 7B | 2024-06-04 | apache-2.0 | Download | |
| Qwen2-7B-Instruct | instruct | 7B | 2024-06-04 | apache-2.0 | Qwen/Qwen2-7B | Download · Qwen2-7B-Instruct-GPTQ-Int4 · Qwen2-7B-Instruct-GPTQ-Int8 · Qwen2-7B-Instruct-AWQ · Qwen2-7B-Instruct-GGUF · Qwen2-7B-Instruct-MLX |
| Qwen2-Audio-7B | base | 7B | 2024-07-16 | apache-2.0 | Download | |
| Qwen2-Audio-7B-Instruct | instruct | 7B | 2024-07-31 | apache-2.0 | Download | |
| Qwen2-Math-1.5B | base | 1.5B | 2024-08-08 | apache-2.0 | Download | |
| Qwen2-Math-1.5B-Instruct | instruct | 1.5B | 2024-08-08 | apache-2.0 | Download | |
| Qwen2-Math-72B | base | 72B | 2024-08-08 | other: tongyi-qianwen | Download | |
| Qwen2-Math-72B-Instruct | instruct | 72B | 2024-08-08 | other: tongyi-qianwen | Download | |
| Qwen2-Math-7B | base | 7B | 2024-08-08 | apache-2.0 | Download | |
| Qwen2-Math-7B-Instruct | instruct | 7B | 2024-08-08 | apache-2.0 | Download | |
| Qwen2-VL-2B-Instruct | instruct | 2B | 2024-08-28 | apache-2.0 | Qwen/Qwen2-VL-2B | Download · Qwen2-VL-2B-Instruct-AWQ · Qwen2-VL-2B-Instruct-GPTQ-Int4 · Qwen2-VL-2B-Instruct-GPTQ-Int8 |
| Qwen2-VL-7B-Instruct | instruct | 7B | 2024-08-28 | apache-2.0 | Qwen/Qwen2-VL-7B | Download · Qwen2-VL-7B-Instruct-AWQ · Qwen2-VL-7B-Instruct-GPTQ-Int4 · Qwen2-VL-7B-Instruct-GPTQ-Int8 |
| Qwen2-VL-2B | base | 2B | 2024-09-05 | apache-2.0 | Download | |
| Qwen2-VL-7B | base | 7B | 2024-09-05 | apache-2.0 | Download | |
| Qwen2-Math-RM-72B | reward model | 72B | 2024-09-17 | other: tongyi-qianwen | Qwen/Qwen2-Math-72B-Instruct | Download |
| Qwen2-VL-72B-Instruct | instruct | 72B | 2024-09-17 | other: tongyi-qianwen | Qwen/Qwen2-VL-72B | Download · Qwen2-VL-72B-Instruct-AWQ · Qwen2-VL-72B-Instruct-GPTQ-Int4 · Qwen2-VL-72B-Instruct-GPTQ-Int8 |
| Qwen2-VL-72B | base | 72B | 2024-12-04 | other: qwen | Download | |
| QVQ-72B-Preview | reasoning | 72B | 2024-12-24 | other: qwen | Qwen/Qwen2-VL-72B | Download |
Qwen1.5 from 2024-01-22
Includes CodeQwen1.5-7B. "released" is the Hub repo creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen1.5-0.5B | base | 0.5B | 2024-01-22 | other: tongyi-qianwen-research | Download | |
| Qwen1.5-1.8B | base | 1.8B | 2024-01-22 | other: tongyi-qianwen-research | Download | |
| Qwen1.5-14B | base | 14B | 2024-01-22 | other: tongyi-qianwen | Download | |
| Qwen1.5-4B | base | 4B | 2024-01-22 | other: tongyi-qianwen-research | Download | |
| Qwen1.5-7B | base | 7B | 2024-01-22 | other: tongyi-qianwen | Download | |
| Qwen1.5-72B | base | 72B | 2024-01-23 | other: tongyi-qianwen | Download | |
| Qwen1.5-1.8B-Chat | instruct | 1.8B | 2024-01-30 | other: tongyi-qianwen-research | Download · Qwen1.5-1.8B-Chat-AWQ · Qwen1.5-1.8B-Chat-GGUF · Qwen1.5-1.8B-Chat-GPTQ-Int8 · Qwen1.5-1.8B-Chat-GPTQ-Int4 | |
| Qwen1.5-14B-Chat | instruct | 14B | 2024-01-30 | other: tongyi-qianwen | Download · Qwen1.5-14B-Chat-AWQ · Qwen1.5-14B-Chat-GGUF · Qwen1.5-14B-Chat-GPTQ-Int8 · Qwen1.5-14B-Chat-GPTQ-Int4 | |
| Qwen1.5-4B-Chat | instruct | 4B | 2024-01-30 | other: tongyi-qianwen-research | Download · Qwen1.5-4B-Chat-AWQ · Qwen1.5-4B-Chat-GGUF · Qwen1.5-4B-Chat-GPTQ-Int8 · Qwen1.5-4B-Chat-GPTQ-Int4 | |
| Qwen1.5-72B-Chat | instruct | 72B | 2024-01-30 | other: tongyi-qianwen | Qwen/Qwen1.5-72B | Download · Qwen1.5-72B-Chat-AWQ · Qwen1.5-72B-Chat-GGUF · Qwen1.5-72B-Chat-GPTQ-Int8 · Qwen1.5-72B-Chat-GPTQ-Int4 |
| Qwen1.5-7B-Chat | instruct | 7B | 2024-01-30 | other: tongyi-qianwen | Download · Qwen1.5-7B-Chat-AWQ · Qwen1.5-7B-Chat-GGUF · Qwen1.5-7B-Chat-GPTQ-Int8 · Qwen1.5-7B-Chat-GPTQ-Int4 | |
| Qwen1.5-0.5B-Chat | instruct | 0.5B | 2024-01-31 | other: tongyi-qianwen-research | Download · Qwen1.5-0.5B-Chat-AWQ · Qwen1.5-0.5B-Chat-GGUF · Qwen1.5-0.5B-Chat-GPTQ-Int4 · Qwen1.5-0.5B-Chat-GPTQ-Int8 | |
| Qwen1.5-MoE-A2.7B | base | 2.7B | 2024-02-29 | other: tongyi-qianwen | Download | |
| Qwen1.5-MoE-A2.7B-Chat | instruct | 2.7B | 2024-03-14 | other: tongyi-qianwen | Download · Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4 | |
| Qwen1.5-32B | base | 32B | 2024-04-01 | other: tongyi-qianwen-research | Download | |
| Qwen1.5-32B-Chat | instruct | 32B | 2024-04-03 | other: tongyi-qianwen | Download · Qwen1.5-32B-Chat-GPTQ-Int4 · Qwen1.5-32B-Chat-GGUF · Qwen1.5-32B-Chat-AWQ | |
| CodeQwen1.5-7B | base | 7B | 2024-04-15 | other: tongyi-qianwen-research | Download · CodeQwen1.5-7B-AWQ | |
| CodeQwen1.5-7B-Chat | instruct | 7B | 2024-04-15 | other: tongyi-qianwen | Download · CodeQwen1.5-7B-Chat-AWQ · CodeQwen1.5-7B-Chat-GGUF | |
| Qwen1.5-110B | base | 110B | 2024-04-25 | other: tongyi-qianwen | Download | |
| Qwen1.5-110B-Chat | instruct | 110B | 2024-04-25 | other: tongyi-qianwen | Download · Qwen1.5-110B-Chat-GPTQ-Int4 · Qwen1.5-110B-Chat-AWQ · Qwen1.5-110B-Chat-GGUF |
Qwen from 2023-08-03
The original Qwen (v1) generation, 2023. Repos without a suffix are base models; -Chat are instruct. "released" is the Hub repo creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen-7B | base | 7B | 2023-08-03 | other: tongyi-qianwen-license-agreement | Download | |
| Qwen-7B-Chat | instruct | 7B | 2023-08-03 | other: tongyi-qianwen-license-agreement | Download · Qwen-7B-Chat-Int4 · Qwen-7B-Chat-Int8 | |
| Qwen-VL | base | not stated | 2023-08-18 | not stated on card | Download | |
| Qwen-VL-Chat | instruct | not stated | 2023-08-20 | not stated on card | Download · Qwen-VL-Chat-Int4 | |
| Qwen-14B | base | 14B | 2023-09-24 | not stated on card | Download | |
| Qwen-14B-Chat | instruct | 14B | 2023-09-24 | not stated on card | Download · Qwen-14B-Chat-Int4 · Qwen-14B-Chat-Int8 | |
| Qwen-72B | base | 72B | 2023-11-26 | other: tongyi-qianwen-license-agreement | Download | |
| Qwen-72B-Chat | instruct | 72B | 2023-11-29 | other: tongyi-qianwen | Qwen/Qwen-72B | Download · Qwen-72B-Chat-Int4 · Qwen-72B-Chat-Int8 |
| Qwen-1_8B | base | 1.8B | 2023-11-30 | not stated on card | Download | |
| Qwen-1_8B-Chat-Int4 | instruct | 1.8B | 2023-11-30 | not stated on card | Download | |
| Qwen-1_8B-Chat-Int8 | instruct | 1.8B | 2023-11-30 | not stated on card | Download | |
| Qwen-Audio | base | not in name; Hub safetensors total 8.4B | 2023-11-30 | not stated on card | Download | |
| Qwen-Audio-Chat | instruct | not in name; Hub safetensors total 8.4B | 2023-11-30 | not stated on card | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Qwen3-30B-A3B-Instruct-2507 5, Qwen3-4B-Instruct-2507 5, Qwen3.5-0.8B 5, Qwen3.5-122B-A10B 5, Qwen3.5-2B 5, Qwen3.5-397B-A17B 5, Qwen3.5-4B 5, Qwen3.5-9B 5, Qwen3.6-27B 5, Qwen3.6-35B-A3B 5, Qwen3.8-27B 5, Qwen3.8-Flash-Next 5, Qwen3-ASR-1.7B 3, Qwen3-Coder-30B-A3B-Instruct 5, Qwen3-Coder-Next 5, Qwen3-Omni-30B-A3B-Instruct 2, Qwen3-VL-2B-Instruct 5, Qwen3-VL-30B-A3B-Instruct 5, Qwen3-VL-32B-Instruct 5, Qwen3-VL-4B-Instruct 5, Qwen3-VL-8B-Instruct 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Apodex AI (apodex)
Publishes from huggingface.co/apodex. Checked 2026-10-21.
Apodex-1 from 2026-06-07
Apodex 1.0 (June 2026) and 1.1 (August 2026) are one major family by the decimal-is-minor rule. Every repository is Apache-2.0 and declares its Qwen3.5 base in the metadata. The 397B Apodex-1.1 is served by Apodex's API (platform.apodex.ai), with no open weights on the Hub. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Apodex-1.0-mini | reasoning (agentic) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-06-07 | apache-2.0 | Qwen/Qwen3.5-35B-A3B | Download |
| Apodex-1.0-4B-SFT | SFT checkpoint | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-06-07 | apache-2.0 | Qwen/Qwen3.5-4B | Download |
| Apodex-1.0-2B-SFT | SFT checkpoint | 2B (2,274,069,824 by the Hub's safetensors count) | 2026-06-07 | apache-2.0 | Qwen/Qwen3.5-2B | Download |
| Apodex-1.0-0.8B-SFT | SFT checkpoint | 0.8B (873,438,784 by the Hub's safetensors count) | 2026-06-07 | apache-2.0 | Qwen/Qwen3.5-0.8B | Download |
| Apodex-1.1-mini Apodex-1.1-mini is Apodex AI's open-weight mixture-of-experts model for long-horizon agent work with files, data, code and tools, a fine-tune of Qwen3.5-35B-A3B with about 36B parameters, about 3B active per token and a 262,144-token context, under Apache 2.0. | reasoning (agentic, long-horizon) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-08-25 | apache-2.0 | Qwen/Qwen3.5-35B-A3B | Our listing (C) · Download · Apodex-1.1-mini-FP8 · Apodex-1.1-mini-NVFP4 · Apodex-1.1-mini-GPTQ-Int4 · Apodex-1.1-mini-GGUF |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Apodex-1.1-mini 2. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
AutoTrust AI Lab (autotrust)
Publishes from huggingface.co/autotrust. Checked 2026-10-21.
JEV from 2026-09-23
AutoTrust's typed-decision models (yes or no, choice, 0 to 5 score with calibrated probabilities), trained as students of TypeSafe AI's closed Jev 1.13 from SargeDev/jev-distill-corpus-v3. AutoTrust's JEV-27B card disclaims any affiliation with TypeSafe AI; don't present one as the other (see data/tools/typesafe-jev.json). Each size is named by its base. The organisation is unverified; its ModelScope organisation AutoTrust-Labs mirrors it. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| JEV-9B | decision model (text) | 9B (8,953,803,264 by the Hub's safetensors count) | 2026-09-23 | apache-2.0 | Qwen/Qwen3.5-9B | Download |
| JEV-27B (Text-only (Qwen3_5ForCausalLM).) | decision model (text) | 27B (26,895,998,464 by the Hub's safetensors count) | 2026-09-25 | apache-2.0 | Qwen/Qwen3.8-27B | Download |
| JEV-Gemma4-26B-A4B | decision model (text) | 26B total, 4B active (MoE; 25,805,936,206 by the Hub's safetensors count) | 2026-09-29 | apache-2.0 | google/gemma-4-26B-A4B-it | Download |
| JEV-27B-VL JEV-27B-VL is AutoTrust AI Lab's open-weight decision model, a LoRA adapter and decision head on Qwen3.8-27B that returns option probabilities for questions over text and images, trained on labels distilled from TypeSafe's Jev. It is not a TypeSafe product. (No quantised build from AutoTrust. 2,019,618 downloads in 30 days and in all time, three weeks after creation (unverified).) | decision model (text and images) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-30 | apache-2.0 | Qwen/Qwen3.8-27B | Our listing (C) · Download |
| JEV-35B | decision model (text) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-10-11 | apache-2.0 | Qwen/Qwen3.5-35B-A3B | Download · JEV-35B-NVFP4 |
| AutoTrust-JEV-27B-v1.1 | decision model | 27B (27,356,728,560 by the Hub's safetensors count) | 2026-10-17 | apache-2.0 | not declared (a Qwen3.5-architecture 27B: Qwen3_5ForConditionalGeneration with 27,356,728,560 parameters, the count other Qwen3.x-27B derivatives show without the MTP layer, so the Qwen generation cannot be told from the files) | Download |
Older versions (1 family, 1 models)
GEV from 2026-10-02
A parallel decision-model line on Gemma 4, not an earlier version of JEV.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GEV-26B-Decide | decision model | 26B (25,805,936,206 by the Hub's safetensors count) | 2026-10-02 | apache-2.0 | google/gemma-4-26B-A4B-it | Download · GEV-26B-Decide-NVFP4 |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: JEV-27B-VL 4. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
BottleCap AI (bottlecapai)
Publishes from huggingface.co/bottlecapai. Checked 2026-10-21.
ThinkingCap from 2026-07-06
BottleCap's ThinkingCap line, fine-tunes that shorten reasoning traces, named by the base they tune. No version number, so one family. The Qwen3.6 edition is Apache-2.0; the Qwen3.8 edition moved to PolyForm Small Business 1.0.0 with a personal-use grant. The bottlecapai organisation is verified. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ThinkingCap-Qwen3.6-27B (Gated (auto approval): the files open after accepting the card's conditions. The GGUF repository is older (1 July 2026) than the BF16 repository (6 July).) | reasoning (shorter thinking) | 27B (27,356,728,560 by the Hub's safetensors count) | 2026-07-06 | apache-2.0 | Qwen/Qwen3.6-27B | Download · ThinkingCap-Qwen3.6-27B-GGUF · ThinkingCap-Qwen3.6-27B-FP8 · ThinkingCap-Qwen3.6-27B-NVFP4 |
| ThinkingCap-Qwen3.8-27B ThinkingCap-Qwen3.8-27B is BottleCap AI's fine-tune of Alibaba's Qwen3.8-27B that shortens its reasoning traces, a 27B dense vision-language model on Hugging Face under PolyForm Small Business 1.0.0, with a gated main repository. (Gated (auto approval): the files open after accepting the card's conditions.) | reasoning (shorter thinking) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-23 | other (polyform-small-business-1.0.0) | Qwen/Qwen3.8-27B | Our listing (C) · Download · ThinkingCap-Qwen3.8-27B-FP8 · ThinkingCap-Qwen3.8-27B-GGUF · ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw · ThinkingCap-Qwen3.8-27B-NVFP4 · ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ · ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: ThinkingCap (Qwen3.8 27B) 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Cohere (Cohere Labs)
Publishes from huggingface.co/CohereLabs. Checked 2026-10-21.
Cohere Transcribe from 2026-03-24
Cohere's open speech recogniser: Transcribe 03-2026 (2.07B by the Hub count) and an Arabic fine-tune (07-2026), both Apache 2.0 and gated with automatic approval. Not a numbered generation, so "major" is false. Cohere's Audio Transcriptions API serves the model, but no Cohere transcription listing exists in data/tools. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| cohere-transcribe-03-2026 Cohere Transcribe 03-2026 is Cohere's open-weight speech recognition model, a 2B Conformer encoder with a light Transformer decoder trained from scratch, transcribing 14 languages from a pre-specified language, under Apache 2.0, in a gated Hugging Face repository. | speech-to-text | 2.07B (Hub safetensors count; the name states none) | 2026-03-24 | apache-2.0 | Our listing (C) · Download (gated) | |
| cohere-transcribe-arabic-07-2026 | speech-to-text (Arabic) | 2.07B (Hub safetensors count; the name states none) | 2026-06-18 | apache-2.0 | Download (gated) |
Older versions (9 families, 29 models)
North from 2026-06-05
Cohere Labs' other open-weight lines, newest first. Most text models before 2026 are cc-by-nc-4.0 and gated with automatic approval; North, command-a-plus-05-2026 and Transcribe are Apache 2.0.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| North-Small-Translate-1.0 | translation | 218.25B (Hub safetensors count; the name states none) | 2026-08-14 | cc-by-nc-4.0 | Download (gated) · North-Small-Translate-1.0-fp8 · North-Small-Translate-1.0-w4a16 | |
| North-Micro-Vision-Instruct | instruct (vision-language) | 2.48B (Hub safetensors count; the name states none) | 2026-08-10 | apache-2.0 | Download | |
| North-Mini-Code-1.0 | instruct (code) | 30.48B (Hub safetensors count; the name states none) | 2026-06-05 | apache-2.0 | Download · North-Mini-Code-1.0-fp8 · North-Mini-Code-1.0-w4a16 |
Tiny Aya from 2026-02-13
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| tiny-aya-base-32K | base (32K context, per name) | 3.35B (Hub safetensors count; the name states none) | 2026-09-08 | cc-by-nc-4.0 | Download (gated) | |
| tiny-aya-l2-thinker | reasoning (per name; card not read) | 3.35B (Hub safetensors count; the name states none) | 2026-09-02 | cc-by-nc-4.0 | Download (gated) | |
| tiny-aya-en-thinker | reasoning (per name; card not read) | 3.35B (Hub safetensors count; the name states none) | 2026-09-02 | cc-by-nc-4.0 | Download (gated) | |
| tiny-aya-water | instruct | 3.35B (Hub safetensors count; the name states none) | 2026-02-13 | cc-by-nc-4.0 | Download (gated) · tiny-aya-water-GGUF | |
| tiny-aya-global | instruct | 3.35B (Hub safetensors count; the name states none) | 2026-02-13 | cc-by-nc-4.0 | Download (gated) · tiny-aya-global-GGUF | |
| tiny-aya-fire | instruct | 3.35B (Hub safetensors count; the name states none) | 2026-02-13 | cc-by-nc-4.0 | Download (gated) · tiny-aya-fire-GGUF | |
| tiny-aya-earth | instruct | 3.35B (Hub safetensors count; the name states none) | 2026-02-13 | cc-by-nc-4.0 | Download (gated) · tiny-aya-earth-GGUF | |
| tiny-aya-base | base | 3.35B (Hub safetensors count; the name states none) | 2026-02-13 | cc-by-nc-4.0 | Download (gated) |
Command A (2025 and 2026) from 2025-03-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| command-a-plus-05-2026 | instruct (vision-language; repository holds the BF16 weights) | 218.75B (Hub safetensors count; the name states none) | 2026-05-11 | apache-2.0 | Download · command-a-plus-05-2026-fp8 · command-a-plus-05-2026-w4a4 | |
| command-a-translate-08-2025 | translation | 111.06B (Hub safetensors count; the name states none) | 2025-08-27 | cc-by-nc-4.0 | Download (gated) | |
| command-a-reasoning-08-2025 | reasoning | 111.06B (Hub safetensors count; the name states none) | 2025-08-12 | cc-by-nc-4.0 | Download (gated) | |
| command-a-vision-07-2025 | instruct (vision-language) | 111.87B (Hub safetensors count; the name states none) | 2025-07-28 | cc-by-nc-4.0 | Download (gated) | |
| c4ai-command-a-03-2025 | instruct | 111.06B (Hub safetensors count; the name states none) | 2025-03-11 | cc-by-nc-4.0 | Download (gated) |
Aya Vision from 2025-03-02
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| aya-vision-8b | instruct (vision-language) | 8B (Hub safetensors count 8.63B) | 2025-03-02 | cc-by-nc-4.0 | Download (gated) | |
| aya-vision-32b | instruct (vision-language) | 32B (Hub safetensors count 33.14B) | 2025-03-02 | cc-by-nc-4.0 | Download (gated) |
Command R7B from 2024-12-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| c4ai-command-r7b-arabic-02-2025 | instruct (Arabic) | 8.03B (Hub safetensors count; the name states none) | 2025-02-27 | cc-by-nc-4.0 | Download (gated) | |
| c4ai-command-r7b-12-2024 | instruct | 8.03B (Hub safetensors count; the name states none) | 2024-12-11 | cc-by-nc-4.0 | Download (gated) |
Aya Expanse from 2024-10-23
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| aya-expanse-8b | instruct | 8B (Hub safetensors count 8.03B) | 2024-10-23 | cc-by-nc-4.0 | Download (gated) | |
| aya-expanse-32b | instruct | 32B (Hub safetensors count 32.30B) | 2024-10-23 | cc-by-nc-4.0 | Download (gated) |
Aya 23 from 2024-05-19
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| aya-23-8B | instruct | 8B (Hub safetensors count 8.03B) | 2024-05-19 | cc-by-nc-4.0 | Download (gated) | |
| aya-23-35B | instruct | 35B (Hub safetensors count 34.98B) | 2024-05-19 | cc-by-nc-4.0 | Download (gated) |
Command R and Command R+ from 2024-03-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| c4ai-command-r-plus-08-2024 | instruct | 103.81B (Hub safetensors count; the name states none) | 2024-08-21 | cc-by-nc-4.0 | Download (gated) | |
| c4ai-command-r-08-2024 | instruct | 32.30B (Hub safetensors count; the name states none) | 2024-08-19 | cc-by-nc-4.0 | Download (gated) | |
| c4ai-command-r-plus | instruct | 103.81B (Hub safetensors count; the name states none) | 2024-04-03 | cc-by-nc-4.0 | Download (gated) · c4ai-command-r-plus-4bit | |
| c4ai-command-r-v01 | instruct | 34.98B (Hub safetensors count; the name states none) | 2024-03-11 | cc-by-nc-4.0 | Download (gated) · c4ai-command-r-v01-4bit |
Aya 101 from 2024-02-08
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| aya-101 | instruct (multilingual, mT5-based) | 12.92B (Hub safetensors count; the name states none) | 2024-02-08 | apache-2.0 | Download |
Conway Research
Publishes from huggingface.co/ConwayResearch. Checked 2026-10-21.
Underdog Bark from 2026-10-15
Speech recognition models from Conway Research, the makers of Underdog. Bark 3.0 is a post-trained and compressed NVIDIA Parakeet TDT 0.6B v3, through moondream/parakeet-ultra, by its NOTICE file. Earlier Bark versions weren't found on the Hub.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Underdog-Bark-3.0 Underdog Bark 3.0 is Conway Research's open-weight English speech recognition model, a post-trained and compressed Parakeet TDT 0.6B v3 in one 163 MB file, run with Conway's loader on Transformers, MLX or Core ML, under CC BY 4.0. | speech to text | 0.6B | 2026-10-15 | cc-by-4.0 | nvidia/parakeet-tdt-0.6b-v3 | Our listing (D) · Download |
DeepSeek
Publishes from huggingface.co/deepseek-ai. Checked 2026-10-21.
DeepSeek V4 from 2026-04-22
Preview (V4-Flash, V4-Pro and their bases) on 2026-04-22, DSpark copies 2026-06-27, official releases V4-Flash-0731 and V4-Pro-0813, an experimental vision model, then V4.1-Flash (2026-09-10) and its 1017 update. DeepSeek calls V4.1 a minor version and the brief puts it in V4, but the V4.1-Flash card says it is a new 552B Causal Encoder-Decoder model trained from scratch, so it is a new architecture under the V4 name. Release dates throughout are the Hugging Face repository creation dates from the Hub API. DeepSeek ships FP8 / FP4+FP8 weights natively and publishes no separate quantised builds in its organisation, so every builds list is empty.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-V4-Flash-Base (No model card in the repository; licence from its LICENSE file (MIT), params from the V4 preview card's download table) | base | 284B total, 13B activated | 2026-04-22 | mit | Download | |
| DeepSeek-V4-Flash (Preview release of V4 Flash; FP4 experts + FP8 mixed precision) | instruct | 284B total, 13B activated | 2026-04-22 | mit | Download | |
| DeepSeek-V4-Flash-DSpark (Card: not a new model, the same checkpoint as V4-Flash with a DSpark speculative-decoding module attached) | instruct | 284B total, 13B activated | 2026-06-27 | mit | deepseek-ai/DeepSeek-V4-Flash | Download |
| DeepSeek-V4-Flash-0731 DeepSeek-V4-Flash-0731 is the 31 July 2026 release of DeepSeek's V4 Flash, a mixture-of-experts text model with 284B parameters, 13B active and a 1M-token context, under MIT. DeepSeek's API retired V4 Flash on 10 September 2026. (Card: the official release of V4-Flash, superseding the preview, same structure as V4-Flash-DSpark (speculative-decoding module attached); reasoning_effort low/high/max) | instruct | 284B total, 13B activated (V4-Flash figures; the 0731 card does not restate them) | 2026-07-31 | mit | Our listing (B) · Download | |
| DeepSeek-V4-Flash-Vision-Exp (Card: first experimental multimodal model in the V4 family, built on the V4-Flash architecture with visual modules and continued training; Hub safetensors count 304.6B) | instruct (multimodal, experimental) | not stated on card | 2026-08-31 | mit | Download | |
| DeepSeek-V4-Pro-Base (No model card in the repository; licence from its LICENSE file (MIT), params from the V4 preview card's download table) | base | 1.6T total, 49B activated | 2026-04-22 | mit | Download | |
| DeepSeek-V4-Pro (Preview release of V4 Pro; FP4 experts + FP8 mixed precision) | instruct | 1.6T total, 49B activated | 2026-04-22 | mit | Download | |
| DeepSeek-V4-Pro-DSpark (Card: not a new model, the same checkpoint as V4-Pro with a DSpark speculative-decoding module attached) | instruct | 1.6T total, 49B activated | 2026-06-27 | mit | deepseek-ai/DeepSeek-V4-Pro | Download |
| DeepSeek-V4-Pro-0813 (Card: the official release of V4-Pro, superseding the preview, built on the V4-Pro (Preview) structure with a DSpark module attached) | instruct | 1.6T total, 49B activated (V4-Pro figures; the 0813 card does not restate them) | 2026-08-13 | mit | Download | |
| DeepSeek-V4.1-Flash DeepSeek V4.1 Flash is DeepSeek's open-weight multimodal mixture-of-experts model, 552B backbone parameters with 8B active in prefill and 16B in decode and a 1M-token context, published on Hugging Face under MIT. The 17 October update is graded. (Card: Causal Encoder-Decoder architecture, trained from scratch on 45T multimodal tokens, 1M context. Its base-model table names a DeepSeek-V4.1-Flash-Base, which has no repository in the organisation) | instruct (multimodal) | 552B backbone, 8B activated in prefill / 16B in decode | 2026-09-10 | mit | Our listing (C) · Download | |
| DeepSeek-V4.1-Flash-1017 DeepSeek V4.1 Flash is DeepSeek's open-weight multimodal mixture-of-experts model, 552B backbone parameters with 8B active in prefill and 16B in decode and a 1M-token context, published on Hugging Face under MIT. The 17 October update is graded. (Card: a minor update to DeepSeek-V4.1-Flash. Our listing deepseek-v4-1-flash grades this update and also lists the September repository) | instruct (multimodal) | 552B backbone, 8B / 16B activated (V4.1-Flash figures; not restated) | 2026-10-17 | mit | deepseek-ai/DeepSeek-V4.1-Flash | Our listing (C) · Download |
DeepSeek V3 from 2024-12-25
V3 (Dec 2024), V3-0324, V3.1 and V3.1-Terminus (hybrid thinking), V3.2-Exp (sparse attention), V3.2 and V3.2-Speciale. V3 and V3-Base weights are under the DeepSeek Model License; from V3-0324 on the weights are MIT. The R1 reasoning line is trained from DeepSeek-V3-Base but DeepSeek names it separately, so it is listed under older as its own family.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-V3-Base (Code MIT, weights under the DeepSeek Model License (LICENSE-MODEL); card says commercial use is allowed) | base | 671B total, 37B activated | 2024-12-25 | deepseek-model-license | Download | |
| DeepSeek-V3 (Code MIT, weights under the DeepSeek Model License (LICENSE-MODEL); card says commercial use is allowed) | instruct | 671B total, 37B activated | 2024-12-25 | deepseek-model-license | Download | |
| DeepSeek-V3-0324 (Hub safetensors count 684.5B) | instruct | not stated on card | 2025-03-24 | mit | Download | |
| DeepSeek-V3.1-Base | base | 671B total, 37B activated | 2025-08-19 | mit | Download | |
| DeepSeek-V3.1 | instruct (hybrid thinking) | 671B total, 37B activated | 2025-08-21 | mit | deepseek-ai/DeepSeek-V3.1-Base | Download |
| DeepSeek-V3.1-Terminus (Hub safetensors count 684.5B) | instruct (hybrid thinking) | not stated on card | 2025-09-22 | mit | deepseek-ai/DeepSeek-V3.1-Base | Download |
| DeepSeek-V3.2-Exp-Base (Card is front matter only; Hub safetensors count 685.4B) | base | not stated on card | 2025-09-29 | mit | Download | |
| DeepSeek-V3.2-Exp (Card: experimental, builds on V3.1-Terminus by adding DeepSeek Sparse Attention; Hub safetensors count 685.4B) | instruct (hybrid thinking) | not stated on card (inference config named config_671B_v3.2.json) | 2025-09-29 | mit | deepseek-ai/DeepSeek-V3.2-Exp-Base | Download |
| DeepSeek-V3.2-Speciale (Card: high-compute variant designed exclusively for deep reasoning, no tool calling; Hub safetensors count 685.4B) | reasoning | not stated on card | 2025-11-28 | mit | deepseek-ai/DeepSeek-V3.2-Exp-Base | Download |
| DeepSeek-V3.2 DeepSeek-V3.2 is DeepSeek's open-weight mixture-of-experts reasoning model with DeepSeek Sparse Attention, trained for tool use with a 128K-token context, published in FP8 under MIT. DeepSeek's API stopped serving it on 24 April 2026, moving to V4 Flash. (Hub safetensors count 685.4B) | instruct (thinking and non-thinking) | not stated on card | 2025-12-01 | mit | deepseek-ai/DeepSeek-V3.2-Exp-Base | Our listing (C) · Download |
Older versions (12 families, 72 models)
DeepSeek-R1 from 2025-01-20
Reasoning line built on the V3 generation: R1-Zero and R1 are trained from DeepSeek-V3-Base (card), R1-0528 is a minor upgrade. The distills are DeepSeek's own fine-tunes of Qwen2.5, Llama 3.1/3.3 and Qwen3 models on R1 outputs, published in deepseek-ai, and belong to this family. V3.1 later folded thinking mode into the V3 line.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-R1-Zero | reasoning (RL only, no SFT) | 671B total, 37B activated | 2025-01-20 | mit | DeepSeek-V3-Base | Download |
| DeepSeek-R1 | reasoning | 671B total, 37B activated | 2025-01-20 | mit | DeepSeek-V3-Base | Download |
| DeepSeek-R1-Distill-Qwen-1.5B | reasoning (distilled) | 1.5B (from the name; not stated) | 2025-01-20 | mit (base model originally Apache 2.0, per card) | Qwen/Qwen2.5-Math-1.5B | Download |
| DeepSeek-R1-Distill-Qwen-7B | reasoning (distilled) | 7B (from the name; not stated) | 2025-01-20 | mit (base model originally Apache 2.0, per card) | Qwen/Qwen2.5-Math-7B | Download |
| DeepSeek-R1-Distill-Llama-8B | reasoning (distilled) | 8B (from the name; not stated) | 2025-01-20 | mit (base model originally under the Llama 3.1 licence, per card) | meta-llama/Llama-3.1-8B | Download |
| DeepSeek-R1-Distill-Qwen-14B | reasoning (distilled) | 14B (from the name; not stated) | 2025-01-20 | mit (base model originally Apache 2.0, per card) | Qwen/Qwen2.5-14B | Download |
| DeepSeek-R1-Distill-Qwen-32B | reasoning (distilled) | 32B (from the name; not stated) | 2025-01-20 | mit (base model originally Apache 2.0, per card) | Qwen/Qwen2.5-32B | Download |
| DeepSeek-R1-Distill-Llama-70B | reasoning (distilled) | 70B (from the name; not stated) | 2025-01-20 | mit (base model originally under the Llama 3.3 licence, per card) | meta-llama/Llama-3.3-70B-Instruct | Download |
| DeepSeek-R1-0528 (Card: a minor version upgrade of DeepSeek R1; Hub safetensors count 684.5B) | reasoning | not stated on card | 2025-05-28 | mit | Download | |
| DeepSeek-R1-0528-Qwen3-8B (Card: chain-of-thought distilled from DeepSeek-R1-0528 to post-train Qwen3 8B Base) | reasoning (distilled) | 8B (from the name; not stated) | 2025-05-29 | mit | Qwen3 8B Base | Download |
DeepSeekMath-V2 from 2025-11-27
Side release built on DeepSeek-V3.2-Exp-Base (card); Apache 2.0.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-Math-V2 (Hub safetensors count 685.4B) | reasoning (theorem proving / self-verification) | not stated on card | 2025-11-27 | apache-2.0 | DeepSeek-V3.2-Exp-Base | Download |
DeepSeek-Prover from 2024-08-15
Lean theorem-proving models: V1, V1.5 (7B base/SFT/RL) and V2 (7B and 671B, 2025-04-30).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-Prover-V1 (Hub safetensors count 6.9B) | instruct (theorem proving) | not stated on card | 2024-08-16 | deepseek-model-license | Download | |
| DeepSeek-Prover-V1.5-Base | base | 7B | 2024-08-15 | deepseek-model-license | Download | |
| DeepSeek-Prover-V1.5-SFT | sft | 7B | 2024-08-15 | deepseek-model-license | deepseek-ai/DeepSeek-Prover-V1.5-Base | Download |
| DeepSeek-Prover-V1.5-RL | rl | 7B | 2024-08-15 | deepseek-model-license | deepseek-ai/DeepSeek-Prover-V1.5-SFT | Download |
| DeepSeek-Prover-V2-7B | instruct (theorem proving) | 7B | 2025-04-30 | none tagged (LICENSE file in repo) | DeepSeek-Prover-V1.5-Base | Download |
| DeepSeek-Prover-V2-671B | instruct (theorem proving) | 671B | 2025-04-30 | none tagged (LICENSE file in repo) | DeepSeek-V3-Base | Download |
DeepSeek-V2.5 from 2024-09-05
Merges V2-Chat and Coder-V2-Instruct (card).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-V2.5 (Hub safetensors count 235.7B) | instruct | not stated on card | 2024-09-05 | deepseek-model-license | DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct (card: combines the two) | Download |
| DeepSeek-V2.5-1210 (Hub safetensors count 235.7B) | instruct | not stated on card | 2024-12-10 | deepseek-model-license | deepseek-ai/DeepSeek-V2.5 | Download |
ESFT (Expert-Specialized Fine-Tuning) from 2024-07-04
Research checkpoints from the ESFT paper: a vanilla model plus 12 task-specific expert sets. No licence stated in the repositories.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ESFT-vanilla-lite (Card: the vanilla model used in the ESFT paper (arXiv 2407.01906); no licence file or tag; Hub safetensors count 15.7B) | research checkpoint | not stated on card | 2024-07-04 | none stated | Download | |
| ESFT-gate-intent-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 2.1B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-intent-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.6B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-gate-code-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 2.3B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-code-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.9B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-gate-law-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.7B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-law-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.5B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-gate-math-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.6B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-math-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.4B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-gate-translation-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.1B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-translation-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 0.9B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-gate-summary-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.7B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
| ESFT-token-summary-lite (Card: a specialised expert set of ESFT-vanilla-lite; Hub safetensors count 1.2B) | expert set (task-specialised) | not stated on card | 2024-07-04 | none stated | deepseek-ai/ESFT-vanilla-lite | Download |
DeepSeek-Coder-V2 from 2024-06-14
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-Coder-V2-Base | base | 236B total, 21B activated | 2024-06-14 | deepseek-model-license | intermediate checkpoint of DeepSeek-V2 (card: further pre-trained from it) | Download |
| DeepSeek-Coder-V2-Instruct | instruct | 236B total, 21B activated | 2024-06-14 | deepseek-model-license | deepseek-ai/DeepSeek-Coder-V2-Base | Download |
| DeepSeek-Coder-V2-Instruct-0724 | instruct | 236B total, 21B activated | 2024-09-05 | deepseek-model-license | deepseek-ai/DeepSeek-Coder-V2-Base | Download |
| DeepSeek-Coder-V2-Lite-Base | base | 16B total, 2.4B activated | 2024-06-14 | deepseek-model-license | intermediate checkpoint of DeepSeek-V2 (card: further pre-trained from it) | Download |
| DeepSeek-Coder-V2-Lite-Instruct | instruct | 16B total, 2.4B activated | 2024-06-14 | deepseek-model-license | Download |
DeepSeek-V2 from 2024-04-22
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DeepSeek-V2 | base | 236B total, 21B activated | 2024-04-22 | deepseek-model-license | Download | |
| DeepSeek-V2-Chat | instruct (RL) | 236B total, 21B activated | 2024-04-28 | deepseek-model-license | Download | |
| DeepSeek-V2-Chat-0628 (Hub safetensors count 235.7B) | instruct | not stated on card | 2024-07-18 | deepseek-model-license | Download | |
| DeepSeek-V2-Lite | base | 16B total, 2.4B activated | 2024-05-15 | deepseek-model-license | Download | |
| DeepSeek-V2-Lite-Chat | instruct (SFT) | 16B total, 2.4B activated | 2024-05-15 | deepseek-model-license | Download |
DeepSeekMath from 2024-02-05
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| deepseek-math-7b-base | base | 7B (from the name) | 2024-02-05 | deepseek-model-license | Download | |
| deepseek-math-7b-instruct | instruct | 7B (from the name) | 2024-02-05 | deepseek-model-license | Download | |
| deepseek-math-7b-rl | rl | 7B (from the name) | 2024-02-05 | deepseek-model-license | Download |
DeepSeekMoE from 2024-01-08
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| deepseek-moe-16b-base | base | 16B (from the name) | 2024-01-08 | deepseek-model-license | Download | |
| deepseek-moe-16b-chat | instruct | 16B (from the name) | 2024-01-09 | deepseek-model-license | Download |
DeepSeek LLM from 2023-11-29
The first DeepSeek general models (V1).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| deepseek-llm-7b-base | base | 7B | 2023-11-29 | deepseek-model-license | Download | |
| deepseek-llm-7b-chat | instruct | 7B | 2023-11-29 | deepseek-model-license | Download | |
| deepseek-llm-67b-base | base | 67B | 2023-11-29 | deepseek-model-license | Download | |
| deepseek-llm-67b-chat | instruct | 67B | 2023-11-29 | deepseek-model-license | Download |
DeepSeek Coder from 2023-10-23
Includes the v1.5 7B models of 2024-01-25.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| deepseek-coder-1.3b-base | base | 1.3B | 2023-10-28 | deepseek-model-license | Download | |
| deepseek-coder-1.3b-instruct | instruct | 1.3B | 2023-10-29 | deepseek-model-license | Download | |
| deepseek-coder-5.7bmqa-base | base | 5.7B | 2023-10-31 | deepseek-model-license | Download | |
| deepseek-coder-6.7b-base | base | 6.7B | 2023-10-23 | deepseek-model-license | Download | |
| deepseek-coder-6.7b-instruct | instruct | 6.7B | 2023-10-29 | deepseek-model-license | Download | |
| deepseek-coder-33b-base | base | 33B | 2023-10-28 | deepseek-model-license | Download | |
| deepseek-coder-33b-instruct | instruct | 33B | 2023-11-01 | deepseek-model-license | Download | |
| deepseek-coder-7b-base-v1.5 | base | 7B (from the name) | 2024-01-25 | deepseek-model-license | DeepSeek-LLM 7B (card: continue pre-trained from it) | Download |
| deepseek-coder-7b-instruct-v1.5 | instruct | 7B (from the name) | 2024-01-25 | deepseek-model-license | Download |
Speculative-decoding drafters (DSpark, DFlash, EAGLE-3) from 2026-06-28
Not an older version and not DeepSeek base models: small draft models (architecture Qwen3DSparkModel etc.) for Qwen3 4B/8B/14B and Gemma 4 12B target models, released alongside the V4 DSpark work. No model cards. Listed here so the organisation inventory is complete.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| dspark_qwen3_4b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 1.4B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dspark_qwen3_8b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 2.4B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dspark_qwen3_14b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 3.4B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dspark_gemma4_12b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 3.4B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dflash_qwen3_4b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 1.3B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dflash_qwen3_8b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 2.3B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dflash_qwen3_14b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 3.3B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| dflash_gemma4_12b_block7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 3.3B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| eagle3_qwen3_4b_ttt7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 0.9B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| eagle3_qwen3_8b_ttt7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 1.5B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| eagle3_qwen3_14b_ttt7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 2.1B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download | |
| eagle3_gemma4_12b_ttt7 (No README; repository holds config.json and model.safetensors only; Hub safetensors count 2.4B) | speculative-decoding draft model | not stated (no model card) | 2026-06-28 | none stated | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: DeepSeek V4.1 Flash 5, DeepSeek-V4-Flash-0731 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
FINAL-Bench (VIDRAFT)
Publishes from huggingface.co/FINAL-Bench. Checked 2026-10-21.
Darwin from 2026-03-31
VIDRAFT's Darwin family: evolutionary merges ('father' and 'mother' models), fine-tunes and transplants of other labs' models (Qwen3.5, 3.6 and 3.8, Gemma 4, Cohere Command A Plus, Mistral), named by size and specialty. The version numbers (V5, V6, V7, V9) belong to the merge engine, not to the models, so all Darwin repositories are one family. Many cards name their parents only in the text, not in the metadata; those bases are marked 'not declared in the metadata'. VIDRAFT Inc. (vidraft.net) links to this organisation; every commit to Darwin-36B-Opus is by one account, SeaWolf-AI. Member dates are the Hub's createdAt.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Darwin-35B-A3B-Opus | reasoning (merge) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-03-31 | apache-2.0 | Qwen/Qwen3.5-35B-A3B + Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled (not declared in the metadata; the card's text names them as father and mother) | Download · Darwin-35B-A3B-Opus-Q8-GGUF |
| Darwin-9B-Opus | reasoning (merge) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-04-04 | apache-2.0 | Qwen/Qwen3.5-9B + Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled (not declared in the metadata; the card's text names them as father and mother) | Download · Darwin-9B-MFP4 |
| Darwin-31B-Opus | reasoning (merge) | 31B (31,273,088,876 by the Hub's safetensors count) | 2026-04-06 | apache-2.0 | google/gemma-4-31B-it + TeichAI/gemma-4-31B-it-Claude-Opus-Distill (not declared in the metadata; the card's text names them as father and mother) | Download |
| Darwin-4B-Opus | reasoning (merge) | 4B (7,996,156,490 by the Hub's safetensors count) | 2026-04-08 | apache-2.0 | google/gemma-4-E4B-it + arsovskidev/Gemma-4-E4B-Claude-4.6-Opus-Reasoning-Distilled | Download |
| Darwin-2B-webgpu (ONNX files for WebGPU; its metadata declares a quantised copy of Gemma 4 E2B IT, not a Darwin merge.) | in-browser ONNX build (any-to-any) | 2B (Gemma 4 E2B) | 2026-04-09 | apache-2.0 | google/gemma-4-E2B-it | Download |
| Darwin-4B-David (The mother named on the card is DavidAU's uncensored (Heretic) Gemma 4 E4B fine-tune.) | reasoning (merge) | 4B (7,996,156,490 by the Hub's safetensors count) | 2026-04-10 | apache-2.0 | FINAL-Bench/Darwin-4B-Opus + DavidAU/gemma-4-E4B-it-The-DECKARD-Expresso-Universe-HERETIC-UNCENSORED-Thinking (not declared in the metadata; the card's text names them as father and mother) | Download |
| Darwin-4B-Genesis (Declares a Gemma-based parent (Darwin-4B-David) and Qwen/Qwen3.5-4B; the files are Gemma 4 architecture.) | reasoning (cross-architecture merge) | 4B (7,518,069,034 by the Hub's safetensors count) | 2026-04-10 | apache-2.0 | FINAL-Bench/Darwin-4B-David + Qwen/Qwen3.5-4B | Download |
| Darwin-27B-KR | reasoning (Korean) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-04-12 | apache-2.0 | FINAL-Bench/Darwin-27B-Opus + Qwen/Qwen3.5-27B | Download |
| Darwin-27B-Opus (Gated (auto approval): the files open after accepting the card's conditions.) | reasoning (merge) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-04-12 | apache-2.0 | a Qwen3.5-architecture 27B (not declared in the metadata; the README is behind the gate and was not read) | Download |
| Darwin-27B-KR-V2 | reasoning (Korean, text only) | 27B (26,895,998,464 by the Hub's safetensors count) | 2026-04-14 | apache-2.0 | FINAL-Bench/Darwin-27B-KR | Download |
| Darwin-27B-KR-V3 | reasoning (Korean, text only) | 27B (26,895,998,464 by the Hub's safetensors count) | 2026-04-15 | apache-2.0 | FINAL-Bench/Darwin-27B-Opus | Download |
| Darwin-TTS-1.7B-Cross (Declares Qwen3-TTS-12Hz-1.7B-Base and Qwen3-1.7B as bases.) | text-to-speech | 1.7B (2,099,234,881 by the Hub's safetensors count) | 2026-04-15 | apache-2.0 | Qwen/Qwen3-TTS-12Hz-1.7B-Base + Qwen/Qwen3-1.7B | Download |
| Darwin-2B-Opus | reasoning (text only) | 2B (1,881,825,088 by the Hub's safetensors count) | 2026-04-18 | apache-2.0 | Qwen/Qwen3.5-2B | Download · Darwin-2B-Opus-LoRA |
| Darwin-36B-Opus Darwin-36B-Opus is VIDRAFT's open-weight text-only mixture-of-experts reasoning model, about 35B parameters with about 3B active per token, made by merging Qwen3.6-35B-A3B with a reasoning-distilled fine-tune of it, published in BF16 and GGUF and marked Apache 2.0. (Text-only; the vision encoder of Qwen3.6-35B-A3B is not included. No LICENSE file; apache-2.0 in metadata.) | reasoning (merge, text only) | 36B (35B total, 3B active, MoE; 34,660,610,688 by the Hub's safetensors count) | 2026-04-22 | apache-2.0 | Qwen/Qwen3.6-35B-A3B + hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled (not declared in the metadata; the card's text names them as base and reasoning parents; the metadata declared both until 3 June 2026) | Our listing (C) · Download · POCKET-35B-GGUF · POCKET-EN-GGUF · POCKET-KR-GGUF · POCKET-KR-MLX |
| Darwin-28B-NEG | reasoning (text only) | 28B (26,895,998,464 by the Hub's safetensors count) | 2026-04-24 | apache-2.0 | FINAL-Bench/Darwin-28B-Opus | Download |
| Darwin-28B-Opus | reasoning (merge) | 28B (27B-class; 27,781,427,952 by the Hub's safetensors count) | 2026-04-24 | apache-2.0 | Qwen/Qwen3.6-27B + rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled (not declared in the metadata; the card's text names them as father and mother) | Download |
| Darwin-9B-NEG | reasoning (merge) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-04-24 | apache-2.0 | FINAL-Bench/Darwin-9B-Opus (not declared in the metadata; the card's text names it as this model's base) | Download |
| Darwin-28B-KR | reasoning (Korean) | 28B (26,895,998,464 by the Hub's safetensors count) | 2026-04-26 | apache-2.0 | FINAL-Bench/Darwin-28B-Opus + FINAL-Bench/Darwin-27B-KR | Download |
| Darwin-28B-KR-Legal | reasoning (Korean legal) | 28B (26,895,998,464 by the Hub's safetensors count) | 2026-04-26 | apache-2.0 | FINAL-Bench/Darwin-28B-KR | Download |
| Darwin-9B-NEG-x-Negentropy-V8 (One parent is Jackrong's Negentropy-claude-opus-4.7-9B (see jackrong.json).) | reasoning (merge, text only) | 9B (8,953,803,264 by the Hub's safetensors count) | 2026-05-09 | apache-2.0 | FINAL-Bench/Darwin-9B-NEG + Jackrong/Negentropy-claude-opus-4.7-9B | Download |
| Darwin-Mistral-G1 | merge of two Mistral models | 123B-class (Ministral3ForCausalLM; 125,025,988,608 by the Hub's safetensors count) | 2026-05-16 | apache-2.0 | mistralai/Devstral-2-123B-Instruct-2512 + mistralai/Mistral-Medium-3.5-128B | Download |
| Darwin-28B-REASON | reasoning (merge) | 28B (26,895,998,464 by the Hub's safetensors count) | 2026-05-17 | apache-2.0 | FINAL-Bench/Darwin-28B-Opus (not declared in the metadata; the card's text names it as the base model) | Download |
| Darwin-28B-Coder (Its base is declared as VIDraft/Darwin-28B-Opus, in the VIDraft organisation, not FINAL-Bench/Darwin-28B-Opus.) | coding | 28B (26,895,998,464 by the Hub's safetensors count) | 2026-05-20 | apache-2.0 | VIDraft/Darwin-28B-Opus | Download · Darwin-28B-Coder-GGUF |
| Darwin-218B-Delphi | merge (Command A Plus) | 218B (218,750,277,872 by the Hub's safetensors count) | 2026-05-23 | apache-2.0 | FINAL-Bench/Darwin-218B-kr + CohereLabs/command-a-plus-05-2026-bf16 | Download |
| Darwin-218B-Expert-Bio | biomedical (Command A Plus) | 218B (218,750,277,872 by the Hub's safetensors count) | 2026-05-23 | apache-2.0 | CohereLabs/command-a-plus-05-2026-bf16 + FINAL-Bench/Darwin-218B-kr | Download |
| Darwin-218B-kr | Korean (on Cohere Command A Plus) | 218B (218,750,277,872 by the Hub's safetensors count) | 2026-05-23 | apache-2.0 | CohereLabs/command-a-plus-05-2026-bf16 | Download |
| Darwin-218B-Expert | expert (Command A Plus) | 218B (218,750,277,872 by the Hub's safetensors count) | 2026-05-24 | apache-2.0 | FINAL-Bench/Darwin-218B-kr | Download |
| Darwin-60B-DUO (Licence gemma in metadata.) | two-model orchestrator | 60B (two ~30B models together) | 2026-05-27 | gemma | FINAL-Bench/Darwin-28B-Opus + FINAL-Bench/Darwin-31B-Opus (not declared in the metadata; inferred from the card's diagram; the files are a DarwinDuoOrchestrator with sub-folders) | Download |
| Darwin-397B-JGOS | reasoning (expert-transplant merge) | 397B total, 17B active (MoE; 403,397,928,944 by the Hub's safetensors count) | 2026-06-14 | apache-2.0 | Qwen/Qwen3.5-397B-A17B | Download |
| Darwin-398B-JGOS | reasoning (expert-transplant merge) | 398B (MoE; 403,397,928,944 by the Hub's safetensors count) | 2026-06-14 | apache-2.0 | Qwen/Qwen3.5-397B-A17B (not declared in the metadata; the card's text says it is built on Qwen 3.5 397B; same parameter count) | Download |
| Darwin-Chimera-4B-Gen1 | reasoning (Chimera) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-06-19 | apache-2.0 | Qwen/Qwen3-4B | Download |
| Darwin-V9-Chimera-4B (Declares Qwen/Qwen3-4B and google/gemma-4-E4B; the files are Qwen3 architecture.) | reasoning (Chimera) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-06-19 | apache-2.0 | Qwen/Qwen3-4B + google/gemma-4-E4B | Download |
| Darwin-V9-Chimera-4B-SFT (Declares Qwen/Qwen3-4B and google/gemma-4-E4B; the files are Qwen3 architecture.) | reasoning (Chimera, SFT) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-06-19 | apache-2.0 | Qwen/Qwen3-4B + google/gemma-4-E4B | Download |
| Darwin-36B-KR | reasoning (Korean) | 36B (35B total, 3B active, MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-06-23 | apache-2.0 | Qwen/Qwen3.6-35B-A3B | Download |
| Darwin-397B-ZTC | reasoning with ZTC decision readout | 397B (MoE, FP8; 403,586,672,624 by the Hub's safetensors count) | 2026-06-25 | apache-2.0 | Qwen/Qwen3.5-397B-A17B (not declared in the metadata; the card says 'built on Qwen 3.5'; parameter count close to Qwen3.5-397B-A17B) | Download |
| Darwin-4B-Chimera (Licence gemma in metadata.) | reasoning (Chimera) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-07-04 | gemma | not declared (Qwen3ForCausalLM with Qwen3-4B's parameter count; the card describes parents from different families) | Download |
| Darwin-Chimera-4B-MoE-2E-PoC | MoE upcycling proof of concept | 4B (8,735,845,376 by the Hub's safetensors count) | 2026-07-04 | apache-2.0 | Qwen/Qwen3.5-4B (not declared in the metadata; the card says expert 0 is Qwen3.5-4B and expert 1 a reasoning fix) | Download |
| Darwin-V9-Chimera-4B-MoE3 (No README (404) and no licence in metadata.) | reasoning (Chimera, MoE) | 4B (9,403,579,137 by the Hub's safetensors count) | 2026-07-04 | none in the card's metadata | not declared (Qwen3MoeForCausalLM) | Download |
| Darwin-27B-RSI | reasoning (self-improvement rounds, text only) | 27B (26,895,998,464 by the Hub's safetensors count) | 2026-09-27 | apache-2.0 | FINAL-Bench/Darwin-27B-Opus | Download · Darwin-27B-RSI-GGUF |
| Darwin-180B-RSI | reasoning (vision-language) | 180B (MoE; 179,999,981,459 by the Hub's safetensors count) | 2026-09-28 | other (qwen-community-1.0) | Qwen/Qwen3.8-Flash-Next (not declared in the metadata; the card's text names it as parent) | Download |
| Darwin-180B-RSI-R3 | reasoning (vision-language) | 180B (MoE; 179,999,981,459 by the Hub's safetensors count) | 2026-10-01 | other (qwen-community-1.0) | Qwen/Qwen3.8-Flash-Next (not declared in the metadata; the card's text names it as parent, R0) | Download · POCKET-Darwin-180B-GGUF |
| Darwin-27B-ZTC | decision model (no generated tokens) | 27B (25,624,600,064 by the Hub's safetensors count) | 2026-10-08 | apache-2.0 | FINAL-Bench/Darwin-27B-RSI (not declared in the metadata; the card's text names it as the backbone) | Download |
| Darwin-27B-ZTC-v2 | decision model (no generated tokens) | 27B (25,624,600,064 by the Hub's safetensors count) | 2026-10-08 | apache-2.0 | FINAL-Bench/Darwin-27B-RSI (not declared; inferred from Darwin-27B-ZTC) | Download |
| Darwin-27B-ZTC-v5 | decision model (no generated tokens) | 27B (25,624,600,064 by the Hub's safetensors count) | 2026-10-11 | apache-2.0 | FINAL-Bench/Darwin-27B-RSI (not declared; inferred from Darwin-27B-ZTC) | Download |
| Darwin-27B-ZTC-v9 | decision model (no generated tokens) | 27B (25,624,600,064 by the Hub's safetensors count) | 2026-10-17 | apache-2.0 | FINAL-Bench/Darwin-27B-RSI (not declared; inferred from Darwin-27B-ZTC) | Download |
| Darwin-35B-ZTC-Vision | vision decision model (no generated tokens) | 35B total, about 3B active (MoE; 34,598,622,576 by the Hub's safetensors count) | 2026-10-17 | apache-2.0 | FINAL-Bench/Darwin-35B-A3B-Opus (not declared in the metadata; the card's text says it is built on it) | Download |
Older versions (6 families, 19 models)
VKORE from 2026-10-14
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| VKORE-46M-NPU | base (pretrained from scratch on an inference NPU) | 46M (45,756,928 by the Hub's safetensors count) | 2026-10-14 | apache-2.0 | none (pretrained from scratch, per the card) | Download |
ZTC-Judge from 2026-09-19
A separate, later line: answer-verification probes on frozen Qwen3.5 bases, not an earlier version of Darwin.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ZTC-Judge-27B | answer verification (probe) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-19 | apache-2.0 | Qwen/Qwen3.5-27B (not declared in the metadata; the card's table names it, revision fc05daec18b0) | Download |
| ZTC-Judge-4B | answer verification (probe) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-09-20 | apache-2.0 | Qwen/Qwen3.5-4B (not declared in the metadata; the card's table names it) | Download |
| ZTC-Judge-9B | answer verification (probe) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-09-20 | apache-2.0 | Qwen/Qwen3.5-9B (not declared in the metadata; the card's table names it) | Download |
POCKET from 2026-07-24
FINAL-Bench's on-device GGUF line. The POCKET builds of Darwin models are listed as builds above; these two declare other labs' models as their source.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| POCKET-26B-GGUF (Declares google/gemma-4-26B-A4B-it as its base without saying whether FINAL-Bench changed the weights. 240,136 downloads in 30 days.) | on-device GGUF | 26B total, 4B active (MoE; 25,233,142,046 by the Hub's GGUF count) | 2026-07-24 | apache-2.0 | google/gemma-4-26B-A4B-it | Download |
| POCKET-Qwen3.8-Flash-Next-GGUF (Declares Qwen/Qwen3.8-Flash-Next as its base; licence qwen-community-1.0.) | on-device GGUF | 180B (MoE; 176,943,899,520 by the Hub's GGUF count) | 2026-09-15 | other (qwen-community-1.0) | Qwen/Qwen3.8-Flash-Next | Download |
Aether from 2026-07-17
FINAL-Bench's own architecture (AETHER), trained from scratch on 144.2B tokens per the card, not built on another lab's model.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Aether-7B-5Attn | base (pretrained from scratch) | 7B (6,589,532,311 by the Hub's safetensors count) | 2026-07-17 | apache-2.0 | none (trained from scratch, per the card) | Download |
| Aether-7B-5Attn-it | instruct | 7B (6,589,532,311 by the Hub's safetensors count) | 2026-07-17 | apache-2.0 | FINAL-Bench/Aether-7B-5Attn | Download |
| AETHER-7B-7Attn-base | base (pretrained) | 7B (6,589,532,311 by the Hub's safetensors count) | 2026-07-20 | apache-2.0 | none declared (own AETHER architecture) | Download |
| Aether-6B-11Attn-base | base (pretrained) | 6B (5,788,481,367 by the Hub's safetensors count) | 2026-07-20 | apache-2.0 | none declared (own AETHER architecture) | Download |
JGOS from 2026-07-04
A separate, later line (Korean reasoning and expert transplants), alongside Darwin-397B/398B-JGOS.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| JGOS-125B-Q4 | reasoning (4-bit) | 125B (MoE; 125,086,497,008 by the Hub's safetensors count) | 2026-07-04 | apache-2.0 | Qwen/Qwen3.5-122B-A10B (not declared; inferred from the licence link to Qwen3.5-122B-A10B and the card's serving commands) | Download |
| JGOS-31B-Citizen-R | reasoning (Korean) | 31B (31,273,086,512 by the Hub's safetensors count) | 2026-07-05 | apache-2.0 | not declared (Gemma 4 architecture, 31B) | Download |
| Ourbox-35B-JGOS | reasoning (Korean, text only) | 35B total, about 3B active (MoE; 34,660,610,688 by the Hub's safetensors count) | 2026-07-10 | apache-2.0 | Qwen/Qwen3.6-35B-A3B (not declared in the metadata; the card's text names it as the backbone) | Download · Ourbox-35B-JGOS-GGUF |
Qwen3.5-2B experiments (lastbrain, Opus-Distill, Opus-SDPO) from 2026-04-17
Small adapters and fine-tunes on Qwen3.5-2B, published the same days as Darwin-2B-Opus.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3.5-2B-Opus-Distill-v1 | experiment (adapter) | 2B | 2026-04-17 | none in the card's metadata | Qwen/Qwen3.5-2B | Download |
| Qwen3.5-2B-Opus-SDPO-v1 | experiment (adapter) | 2B | 2026-04-17 | none in the card's metadata | Qwen/Qwen3.5-2B | Download |
| lastbrain | experiment (adapter) | not stated (1,881,825,088 by the Hub's safetensors count) | 2026-04-17 | apache-2.0 | Qwen/Qwen3.5-2B | Download |
| lastbrain-v2 | experiment (adapter) | not stated | 2026-04-18 | none in the card's metadata | Qwen/Qwen3.5-2B | Download |
| lastbrain-v3 | experiment (adapter) | not stated | 2026-04-18 | apache-2.0 | Qwen/Qwen3.5-2B | Download |
| lastbrain-v3-private | experiment | not stated (1,881,825,088 by the Hub's safetensors count) | 2026-04-18 | apache-2.0 | Qwen/Qwen3.5-2B | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Darwin-36B-Opus 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Google (Gemma)
Publishes from huggingface.co/google. Checked 2026-10-21.
Gemma 4 from 2026-03-31
Google's fourth mainline Gemma generation (E2B, E4B, 12B, 26B A4B, 31B), all under Apache 2.0 and none gated on Hugging Face. gemma.ai's family tree places DiffusionGemma and EmbeddingGemma 2 under Gemma 4. The -assistant repositories are Google's MTP drafter models for speculative decoding, released 16 April 2026 (12B's with the 12B on 3 June 2026). Google's QAT repositories (q4_0 unquantized, Q4_0 GGUF, w4a16 compressed-tensors, mobile) are listed as builds of the model they quantise. DiffusionGemma has no entry on gemma.ai's releases page; its Hugging Face repository was created 2026-06-09.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gemma-4-26B-A4B | base | 25.2B total, 3.8B active | 2026-03-31 | apache-2.0 | Download | |
| gemma-4-26B-A4B-it Gemma 4 26B A4B IT is Google DeepMind's open-weight instruction-tuned mixture-of-experts model, 25.2B parameters with 3.8B active, taking text and images with a 262,144-token context, published on Hugging Face under Apache 2.0. | instruct | 25.2B total, 3.8B active | 2026-03-31 | apache-2.0 | google/gemma-4-26B-A4B | Our listing (BB) · Download · gemma-4-26B-A4B-it-qat-q4_0-unquantized · gemma-4-26B-A4B-it-qat-q4_0-gguf |
| gemma-4-31B | base | 30.7B | 2026-03-31 | apache-2.0 | Download | |
| gemma-4-31B-it Gemma 4 31B IT is Google DeepMind's open-weight dense multimodal model of 30.7B parameters, taking text and images with a 256K-token context, published on Hugging Face in BF16 with 4-bit QAT builds under Apache 2.0. | instruct | 30.7B | 2026-03-31 | apache-2.0 | google/gemma-4-31B | Our listing (BB) · Download · gemma-4-31B-it-qat-q4_0-unquantized · gemma-4-31B-it-qat-q4_0-gguf · gemma-4-31B-it-qat-w4a16-ct |
| gemma-4-E2B | base | 2.3B effective (5.1B with embeddings) | 2026-03-31 | apache-2.0 | Download | |
| gemma-4-E2B-it Gemma 4 E2B IT is Google DeepMind's small open-weight instruction-tuned model for phones and laptops, 2.3B effective parameters (5.1B with per-layer embeddings), taking text, images, video frames and audio with a 131,072-token context, published on Hugging Face under Apache 2.0. | instruct | 2.3B effective (5.1B with embeddings) | 2026-03-31 | apache-2.0 | google/gemma-4-E2B | Our listing (C) · Download · gemma-4-E2B-it-qat-q4_0-unquantized · gemma-4-E2B-it-qat-q4_0-gguf · gemma-4-E2B-it-qat-mobile-ct · gemma-4-E2B-it-qat-mobile-transformers · gemma-4-E2B-it-qat-w4a16-ct |
| gemma-4-E4B | base | 4.5B effective (8B with embeddings) | 2026-03-31 | apache-2.0 | Download | |
| gemma-4-E4B-it Gemma 4 E4B IT is Google DeepMind's small open-weight instruction-tuned model for phones and laptops, 4.5B effective parameters (8B with per-layer embeddings), taking text, images, video frames and audio with a 131,072-token context, published on Hugging Face under Apache 2.0. | instruct | 4.5B effective (8B with embeddings) | 2026-03-31 | apache-2.0 | google/gemma-4-E4B | Our listing (B) · Download · gemma-4-E4B-it-qat-q4_0-unquantized · gemma-4-E4B-it-qat-q4_0-gguf · gemma-4-E4B-it-qat-mobile-ct · gemma-4-E4B-it-qat-mobile-transformers · gemma-4-E4B-it-qat-w4a16-ct |
| gemma-4-26B-A4B-it-assistant | draft (MTP speculative decoding) | 420M (file metadata on Hugging Face; drafter) | 2026-04-16 | apache-2.0 | Download · gemma-4-26B-A4B-it-qat-q4_0-unquantized-assistant | |
| gemma-4-31B-it-assistant | draft (MTP speculative decoding) | 470M (file metadata on Hugging Face; drafter) | 2026-04-16 | apache-2.0 | Download · gemma-4-31B-it-qat-q4_0-unquantized-assistant | |
| gemma-4-E2B-it-assistant | draft (MTP speculative decoding) | 78M (file metadata on Hugging Face; drafter) | 2026-04-16 | apache-2.0 | Download · gemma-4-E2B-it-qat-q4_0-unquantized-assistant | |
| gemma-4-E4B-it-assistant | draft (MTP speculative decoding) | 79M (file metadata on Hugging Face; drafter) | 2026-04-16 | apache-2.0 | Download · gemma-4-E4B-it-qat-q4_0-unquantized-assistant | |
| gemma-4-12B | base | 11.95B | 2026-06-03 | apache-2.0 | Download | |
| gemma-4-12B-it Gemma 4 12B IT is Google DeepMind's open-weight instruction-tuned dense model, 11.95B parameters in an encoder-free design, taking text, images, video frames and audio with a 262,144-token context, published on Hugging Face under Apache 2.0. | instruct | 11.95B | 2026-06-03 | apache-2.0 | google/gemma-4-12B | Our listing (BB) · Download · gemma-4-12B-it-qat-q4_0-unquantized · gemma-4-12B-it-qat-q4_0-gguf · gemma-4-12B-it-qat-w4a16-ct |
| gemma-4-12B-it-assistant | draft (MTP speculative decoding) | 423M (file metadata on Hugging Face; drafter) | 2026-06-03 | apache-2.0 | Download · gemma-4-12B-it-qat-q4_0-unquantized-assistant | |
| diffusiongemma-26B-A4B-it DiffusionGemma 26B A4B IT is Google DeepMind's experimental open-weight text diffusion model on the Gemma 4 26B A4B mixture-of-experts architecture, 25.2B parameters with 3.8B active, taking text, images and video and writing text in 256-token blocks, under Apache 2.0. | instruct | 25.2B total, 3.8B active | apache-2.0 | Our listing (D) · Download | ||
| embeddinggemma-2 | embedding | 740M | 2026-10-06 | apache-2.0 | Download |
Older versions (5 families, 171 models)
Gemma 3 from 2025-03-10
Gemma 3 mainline (270M to 27B) plus the variants Google builds on it: Gemma 3n, ShieldGemma 2, MedGemma, MedGemma 1.5, EmbeddingGemma, FunctionGemma, T5Gemma 2 and TranslateGemma. Placement follows gemma.ai's family tree or the card's base_model where the tree doesn't cover a variant. Repositories are gated: Hugging Face asks you to accept the Gemma Terms of Use (MedGemma: the Health AI Developer Foundations terms) before download.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gemma-3-12b-it | instruct | 12B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-12b-pt | Download (gated) · gemma-3-12b-it-qat-q4_0-gguf · gemma-3-12b-it-qat-q4_0-unquantized · gemma-3-12b-it-qat-int4-unquantized |
| gemma-3-12b-pt | base | 12B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-12b | Download (gated) · gemma-3-12b-pt-qat-q4_0-gguf |
| gemma-3-1b-it | instruct | 1B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-1b-pt | Download (gated) · gemma-3-1b-it-qat-q4_0-gguf · gemma-3-1b-it-qat-q4_0-unquantized · gemma-3-1b-it-qat-int4-unquantized |
| gemma-3-1b-pt | base | 1B | 2025-03-10 | gemma (Gemma Terms of Use) | Download (gated) · gemma-3-1b-pt-qat-q4_0-gguf | |
| gemma-3-27b-it | instruct | 27B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-27b-pt | Download (gated) · gemma-3-27b-it-qat-q4_0-gguf · gemma-3-27b-it-qat-q4_0-unquantized |
| gemma-3-27b-pt | base | 27B | 2025-03-10 | gemma (Gemma Terms of Use) | Download (gated) · gemma-3-27b-pt-qat-q4_0-gguf | |
| gemma-3-4b-it | instruct | 4B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-4b-pt | Download (gated) · gemma-3-4b-it-qat-q4_0-gguf · gemma-3-4b-it-qat-q4_0-unquantized · gemma-3-4b-it-qat-int4-unquantized |
| gemma-3-4b-pt | base | 4B | 2025-03-10 | gemma (Gemma Terms of Use) | Download (gated) · gemma-3-4b-pt-qat-q4_0-gguf | |
| gemma-3-270m | base | 270M | 2025-08-14 | gemma (Gemma Terms of Use) | Download (gated) · gemma-3-270m-qat-q4_0-unquantized | |
| gemma-3-270m-it | instruct | 270M | 2025-08-14 | gemma (Gemma Terms of Use) | google/gemma-3-270m | Download (gated) · gemma-3-270m-it-qat-q4_0-unquantized |
| embeddinggemma-300m EmbeddingGemma 300M is Google DeepMind's open-weight text embedding model of 308M parameters, returning 768-dimension vectors that truncate to 512, 256 or 128 from up to 2,048 tokens in over 100 languages, gated under the Gemma Terms of Use. | embedding | 308M | 2025-09-04 | gemma (Gemma Terms of Use) | Our listing (D) · Download (gated) · embeddinggemma-300m-qat-q4_0-unquantized · embeddinggemma-300m-qat-q8_0-unquantized | |
| functiongemma-270m-it | instruct | 270M | 2025-12-18 | gemma (Gemma Terms of Use) | Download (gated) | |
| gemma-3n-E2B | base | E2B | 2025-06-26 | gemma (Gemma Terms of Use) | google/gemma-3n-E4B | Download (gated) |
| gemma-3n-E2B-it | instruct | E2B | 2025-06-26 | gemma (Gemma Terms of Use) | google/gemma-3n-E4B-it | Download (gated) · gemma-3n-E2B-it-litert-preview · gemma-3n-E2B-it-litert-lm |
| gemma-3n-E4B | base | E4B | 2025-06-26 | gemma (Gemma Terms of Use) | Download (gated) | |
| gemma-3n-E4B-it | instruct | E4B | 2025-06-26 | gemma (Gemma Terms of Use) | google/gemma-3n-E4B | Download (gated) · gemma-3n-E4B-it-litert-preview · gemma-3n-E4B-it-litert-lm |
| medgemma-27b-text-it | instruct | 27B | 2025-05-20 | health-ai-developer-foundations-terms | google/gemma-3-27b-pt | Download (gated) |
| medgemma-4b-it | instruct | 4B | 2025-05-20 | health-ai-developer-foundations-terms | google/medgemma-4b-pt | Download (gated) |
| medgemma-4b-pt | base | 4B | 2025-05-20 | health-ai-developer-foundations-terms | google/gemma-3-4b-pt | Download (gated) |
| medgemma-27b-it | instruct | 27B | 2025-07-09 | health-ai-developer-foundations-terms | google/gemma-3-27b-pt | Download (gated) |
| medgemma-1.5-4b-it | instruct | 4B | 2026-01-13 | health-ai-developer-foundations-terms | Download (gated) | |
| shieldgemma-2-4b-it | safety classifier | 4B | 2025-03-10 | gemma (Gemma Terms of Use) | google/gemma-3-4b-pt | Download (gated) |
| t5gemma-2-1b-1b | base | 1B-1B (encoder-decoder) | 2025-12-18 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-2-270m-270m | base | 270M-270M (encoder-decoder) | 2025-12-18 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-2-4b-4b | base | 4B-4B (encoder-decoder) | 2025-12-18 | gemma (Gemma Terms of Use) | Download (gated) | |
| translategemma-12b-it | instruct | 12B | 2026-01-15 | gemma (Gemma Terms of Use) | Download (gated) | |
| translategemma-27b-it | instruct | 27B | 2026-01-15 | gemma (Gemma Terms of Use) | Download (gated) | |
| translategemma-4b-it | instruct | 4B | 2026-01-15 | gemma (Gemma Terms of Use) | Download (gated) |
Gemma 2 from 2024-06-27
Gemma 2 mainline (2B, 9B, 27B) plus PaliGemma 2, ShieldGemma, DataGemma, Gemma 2 JPN, TxGemma and T5Gemma. All gated behind the Gemma Terms of Use (TxGemma: Health AI Developer Foundations terms).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gemma-2-27b | base | 27B | 2024-06-27 | gemma (Gemma Terms of Use) | Download (gated) · gemma-2-27b-pytorch | |
| gemma-2-27b-it | instruct | 27B | 2024-06-27 | gemma (Gemma Terms of Use) | google/gemma-2-27b | Download (gated) · gemma-2-27b-it-pytorch |
| gemma-2-9b | base | 9B | 2024-06-27 | gemma (Gemma Terms of Use) | Download (gated) · gemma-2-9b-pytorch · gemma-2-9b-keras | |
| gemma-2-9b-it | instruct | 9B | 2024-06-27 | gemma (Gemma Terms of Use) | google/gemma-2-9b | Download (gated) · gemma-2-9b-it-pytorch · gemma-2-instruct-9b-keras |
| gemma-2-2b | base | 2B | 2024-07-31 | gemma (Gemma Terms of Use) | Download (gated) · gemma-2-2b-pytorch · gemma-2-2b-GGUF | |
| gemma-2-2b-it | instruct | 2B | 2024-07-31 | gemma (Gemma Terms of Use) | google/gemma-2-2b | Download (gated) · gemma-2-2b-it-GGUF · gemma-2-2b-it-pytorch |
| datagemma-rag-27b-it | instruct | 27B | 2024-09-12 | gemma (Gemma Terms of Use) | google/gemma-2-27b-it | Download (gated) |
| datagemma-rig-27b-it | instruct | 27B | 2024-09-12 | gemma (Gemma Terms of Use) | google/gemma-2-27b-it | Download (gated) |
| gemma-2-2b-jpn-it | instruct | 2B | 2024-10-03 | gemma (Gemma Terms of Use) | google/gemma-2-2b-it | Download (gated) · gemma-2-2b-jpn-it-pytorch · gemma-2-2b-jpn-it-flax |
| paligemma2-10b-ft-docci-448 | fine-tuned (docci, 448px) | 10B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-ft-docci-448-jax | |
| paligemma2-10b-pt-224 | base | 10B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-pt-224-jax · paligemma2-10b-pt-224-keras | |
| paligemma2-10b-pt-448 | base | 10B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-pt-448-jax · paligemma2-10b-pt-448-keras | |
| paligemma2-10b-pt-896 | base | 10B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-pt-896-jax · paligemma2-10b-pt-896-keras | |
| paligemma2-28b-pt-224 | base | 28B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-28b-pt-224-jax · paligemma2-28b-pt-224-keras | |
| paligemma2-28b-pt-448 | base | 28B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-28b-pt-448-jax · paligemma2-28b-pt-448-keras | |
| paligemma2-28b-pt-896 | base | 28B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-28b-pt-896-jax · paligemma2-28b-pt-896-keras | |
| paligemma2-3b-ft-docci-448 | fine-tuned (docci, 448px) | 3B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-ft-docci-448-jax | |
| paligemma2-3b-pt-224 | base | 3B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-pt-224-jax · paligemma2-3b-pt-224-keras | |
| paligemma2-3b-pt-448 | base | 3B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-pt-448-jax · paligemma2-3b-pt-448-keras | |
| paligemma2-3b-pt-896 | base | 3B | 2024-12-05 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-pt-896-jax · paligemma2-3b-pt-896-keras | |
| paligemma2-10b-mix-224 | instruct (mix) | 10B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-mix-224-jax | |
| paligemma2-10b-mix-448 | instruct (mix) | 10B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-10b-mix-448-jax | |
| paligemma2-28b-mix-224 | instruct (mix) | 28B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-28b-mix-224-jax | |
| paligemma2-28b-mix-448 | instruct (mix) | 28B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-28b-mix-448-jax | |
| paligemma2-3b-mix-224 | instruct (mix) | 3B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-mix-224-jax | |
| paligemma2-3b-mix-448 | instruct (mix) | 3B | 2025-02-19 | gemma (Gemma Terms of Use) | Download (gated) · paligemma2-3b-mix-448-jax | |
| shieldgemma-27b | safety classifier | 27B | 2024-07-31 | gemma (Gemma Terms of Use) | Download (gated) | |
| shieldgemma-2b | safety classifier | 2B | 2024-07-31 | gemma (Gemma Terms of Use) | Download (gated) | |
| shieldgemma-9b | safety classifier | 9B | 2024-07-31 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-2b-2b-prefixlm | base | 2B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-2b-2b-prefixlm-it | instruct | 2B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-2b-2b-prefixlm | Download (gated) |
| t5gemma-2b-2b-ul2 | base | 2B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-2b-2b-ul2-it | instruct | 2B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-2b-2b-ul2 | Download (gated) |
| t5gemma-9b-2b-prefixlm | base | 9B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-9b-2b-prefixlm-it | instruct | 9B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-9b-2b-prefixlm | Download (gated) |
| t5gemma-9b-2b-ul2 | base | 9B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-9b-2b-ul2-it | instruct | 9B-2B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-9b-2b-ul2 | Download (gated) |
| t5gemma-9b-9b-prefixlm | base | 9B-9B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-9b-9b-prefixlm-it | instruct | 9B-9B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-9b-9b-prefixlm | Download (gated) |
| t5gemma-9b-9b-ul2 | base | 9B-9B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-9b-9b-ul2-it | instruct | 9B-9B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-9b-9b-ul2 | Download (gated) |
| t5gemma-b-b-prefixlm | base | B-B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-b-b-prefixlm-it | instruct | B-B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-b-b-prefixlm | Download (gated) |
| t5gemma-b-b-ul2 | base | B-B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-b-b-ul2-it | instruct | B-B (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-b-b-ul2 | Download (gated) |
| t5gemma-l-l-prefixlm | base | L-L (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-l-l-prefixlm-it | instruct | L-L (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-l-l-prefixlm | Download (gated) |
| t5gemma-l-l-ul2 | base | L-L (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-l-l-ul2-it | instruct | L-L (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-l-l-ul2 | Download (gated) |
| t5gemma-ml-ml-prefixlm | base | ML-ML (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-ml-ml-prefixlm-it | instruct | ML-ML (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-ml-ml-prefixlm | Download (gated) |
| t5gemma-ml-ml-ul2 | base | ML-ML (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-ml-ml-ul2-it | instruct | ML-ML (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-ml-ml-ul2 | Download (gated) |
| t5gemma-s-s-prefixlm | base | S-S (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-s-s-prefixlm-it | instruct | S-S (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-s-s-prefixlm | Download (gated) |
| t5gemma-s-s-ul2 | base | S-S (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-s-s-ul2-it | instruct | S-S (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-s-s-ul2 | Download (gated) |
| t5gemma-xl-xl-prefixlm | base | XL-XL (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-xl-xl-prefixlm-it | instruct | XL-XL (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-xl-xl-prefixlm | Download (gated) |
| t5gemma-xl-xl-ul2 | base | XL-XL (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | Download (gated) | |
| t5gemma-xl-xl-ul2-it | instruct | XL-XL (encoder-decoder) | 2025-07-09 | gemma (Gemma Terms of Use) | google/t5gemma-xl-xl-ul2 | Download (gated) |
| txgemma-27b-chat | chat | 27B | health-ai-developer-foundations-terms | Download (gated) | ||
| txgemma-27b-predict | predict | 27B | health-ai-developer-foundations-terms | Download (gated) | ||
| txgemma-2b-predict | predict | 2B | health-ai-developer-foundations-terms | Download (gated) | ||
| txgemma-9b-chat | chat | 9B | health-ai-developer-foundations-terms | Download (gated) | ||
| txgemma-9b-predict | predict | 9B | health-ai-developer-foundations-terms | Download (gated) |
Gemma from 2024-02-21
First generation (2B, 7B), Gemma 1.1, CodeGemma, CodeGemma 1.1, PaliGemma and Gemma APS. All gated behind the Gemma Terms of Use. PaliGemma's task-specific ft-* checkpoints are each listed as a member.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gemma-2b | base | 2B | 2024-02-21 | gemma (Gemma Terms of Use) | Download (gated) · gemma-2b-GGUF · gemma-2b-cpp · gemma-2b-pytorch · gemma-2b-keras · gemma-2b-sfp-cpp · gemma-2b-flax · gemma-2b-AWQ | |
| gemma-2b-it | instruct | 2B | 2024-02-21 | gemma (Gemma Terms of Use) | Download (gated) · gemma-2b-it-GGUF · gemma-2b-it-cpp · gemma-2b-it-pytorch · gemma-2b-it-keras · gemma-2b-it-sfp-cpp · gemma-2b-it-flax · gemma-2b-it-tflite | |
| gemma-7b | base | 7B | 2024-02-21 | gemma (Gemma Terms of Use) | Download (gated) · gemma-7b-GGUF · gemma-7b-cpp · gemma-7b-pytorch · gemma-7b-keras · gemma-7b-sfp-cpp · gemma-7b-quant-pytorch · gemma-7b-flax · gemma-7b-AWQ | |
| gemma-7b-it | instruct | 7B | 2024-02-21 | gemma (Gemma Terms of Use) | google/gemma-7b | Download (gated) · gemma-7b-it-GGUF · gemma-7b-it-cpp · gemma-7b-it-pytorch · gemma-7b-it-keras · gemma-7b-it-sfp-cpp · gemma-7b-it-quant-pytorch · gemma-7b-it-flax |
| codegemma-2b | base | 2B | 2024-04-09 | gemma (Gemma Terms of Use) | Download (gated) · codegemma-2b-GGUF · codegemma-2b-pytorch · codegemma-2b-keras | |
| codegemma-7b | base | 7B | 2024-04-09 | gemma (Gemma Terms of Use) | Download (gated) · codegemma-7b-GGUF · codegemma-7b-pytorch · codegemma-7b-keras | |
| codegemma-7b-it | instruct | 7B | 2024-04-09 | gemma (Gemma Terms of Use) | Download (gated) · codegemma-7b-it-GGUF · codegemma-7b-it-pytorch · codegemma-7b-it-keras | |
| codegemma-1.1-2b | base | 2B | 2024-05-03 | gemma (Gemma Terms of Use) | Download (gated) · codegemma-1.1-2b-GGUF · codegemma-1.1-2b-pytorch · codegemma-1.1-2b-keras | |
| codegemma-1.1-7b-it | instruct | 7B | 2024-05-03 | gemma (Gemma Terms of Use) | Download (gated) · codegemma-1.1-7b-it-GGUF · codegemma-1.1-7b-it-pytorch · codegemma-1.1-7b-it-keras | |
| gemma-1.1-2b-it | instruct | 2B | 2024-04-05 | gemma (Gemma Terms of Use) | Download (gated) · gemma-1.1-2b-it-GGUF · gemma-1.1-2b-it-pytorch · gemma-1.1-2b-it-tflite · gemma-1.1-2b-it-keras | |
| gemma-1.1-7b-it | instruct | 7B | 2024-04-05 | gemma (Gemma Terms of Use) | Download (gated) · gemma-1.1-7b-it-pytorch · gemma-1.1-7b-it-GGUF · gemma-1.1-7b-it-keras | |
| gemma-2b-aps-it | instruct | 2B | 2024-10-15 | gemma (Gemma Terms of Use) | google/gemma-2b | Download (gated) |
| gemma-7b-aps-it | instruct | 7B | 2024-10-15 | gemma (Gemma Terms of Use) | google/gemma-7b | Download (gated) |
| paligemma-3b-ft-ai2d-224 | fine-tuned (ai2d, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-ai2d-224-jax | |
| paligemma-3b-ft-ai2d-448 | fine-tuned (ai2d, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-ai2d-448-jax | |
| paligemma-3b-ft-aokvqa-da-224 | fine-tuned (aokvqa-da, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-aokvqa-da-224-jax | |
| paligemma-3b-ft-aokvqa-da-448 | fine-tuned (aokvqa-da, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-aokvqa-da-448-jax | |
| paligemma-3b-ft-aokvqa-mc-224 | fine-tuned (aokvqa-mc, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-aokvqa-mc-224-jax | |
| paligemma-3b-ft-aokvqa-mc-448 | fine-tuned (aokvqa-mc, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-aokvqa-mc-448-jax | |
| paligemma-3b-ft-coco35l-224 | fine-tuned (coco35l, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-coco35l-224-jax | |
| paligemma-3b-ft-coco35l-448 | fine-tuned (coco35l, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-coco35l-448-jax | |
| paligemma-3b-ft-cococap-224 | fine-tuned (cococap, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-cococap-224-jax | |
| paligemma-3b-ft-cococap-448 | fine-tuned (cococap, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-cococap-448-jax | |
| paligemma-3b-ft-docvqa-224 | fine-tuned (docvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-docvqa-224-jax | |
| paligemma-3b-ft-docvqa-448 | fine-tuned (docvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-docvqa-448-jax | |
| paligemma-3b-ft-docvqa-896 | fine-tuned (docvqa, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-docvqa-896-jax | |
| paligemma-3b-ft-gqa-224 | fine-tuned (gqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-gqa-224-jax | |
| paligemma-3b-ft-gqa-448 | fine-tuned (gqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-gqa-448-jax | |
| paligemma-3b-ft-infovqa-224 | fine-tuned (infovqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-infovqa-224-jax | |
| paligemma-3b-ft-infovqa-448 | fine-tuned (infovqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-infovqa-448-jax | |
| paligemma-3b-ft-infovqa-896 | fine-tuned (infovqa, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-infovqa-896-jax | |
| paligemma-3b-ft-nlvr2-224 | fine-tuned (nlvr2, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-nlvr2-224-jax | |
| paligemma-3b-ft-nlvr2-448 | fine-tuned (nlvr2, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-nlvr2-448-jax | |
| paligemma-3b-ft-ocrvqa-224 | fine-tuned (ocrvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-ocrvqa-224-jax | |
| paligemma-3b-ft-ocrvqa-448 | fine-tuned (ocrvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-ocrvqa-448-jax | |
| paligemma-3b-ft-ocrvqa-896 | fine-tuned (ocrvqa, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-ocrvqa-896-jax | |
| paligemma-3b-ft-okvqa-224 | fine-tuned (okvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-okvqa-224-jax | |
| paligemma-3b-ft-okvqa-448 | fine-tuned (okvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-okvqa-448-jax | |
| paligemma-3b-ft-refcoco-seg-224 | fine-tuned (refcoco-seg, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-refcoco-seg-224-jax | |
| paligemma-3b-ft-refcoco-seg-448 | fine-tuned (refcoco-seg, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-refcoco-seg-448-jax | |
| paligemma-3b-ft-refcoco-seg-896 | fine-tuned (refcoco-seg, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-refcoco-seg-896-jax | |
| paligemma-3b-ft-rsvqa-hr-224 | fine-tuned (rsvqa-hr, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-rsvqa-hr-224-jax | |
| paligemma-3b-ft-rsvqa-hr-448 | fine-tuned (rsvqa-hr, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-rsvqa-hr-448-jax | |
| paligemma-3b-ft-rsvqa-lr-224 | fine-tuned (rsvqa-lr, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-rsvqa-lr-224-jax | |
| paligemma-3b-ft-rsvqa-lr-448 | fine-tuned (rsvqa-lr, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-rsvqa-lr-448-jax | |
| paligemma-3b-ft-scicap-224 | fine-tuned (scicap, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-scicap-224-jax | |
| paligemma-3b-ft-scicap-448 | fine-tuned (scicap, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-scicap-448-jax | |
| paligemma-3b-ft-science-qa-224 | fine-tuned (science-qa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-science-qa-224-jax | |
| paligemma-3b-ft-science-qa-448 | fine-tuned (science-qa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-science-qa-448-jax | |
| paligemma-3b-ft-screen2words-224 | fine-tuned (screen2words, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-screen2words-224-jax | |
| paligemma-3b-ft-screen2words-448 | fine-tuned (screen2words, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-screen2words-448-jax | |
| paligemma-3b-ft-stvqa-224 | fine-tuned (stvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-stvqa-224-jax | |
| paligemma-3b-ft-stvqa-448 | fine-tuned (stvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-stvqa-448-jax | |
| paligemma-3b-ft-stvqa-896 | fine-tuned (stvqa, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-stvqa-896-jax | |
| paligemma-3b-ft-tallyqa-224 | fine-tuned (tallyqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-tallyqa-224-jax | |
| paligemma-3b-ft-tallyqa-448 | fine-tuned (tallyqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-tallyqa-448-jax | |
| paligemma-3b-ft-textcaps-224 | fine-tuned (textcaps, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-textcaps-224-jax | |
| paligemma-3b-ft-textcaps-448 | fine-tuned (textcaps, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-textcaps-448-jax | |
| paligemma-3b-ft-textvqa-224 | fine-tuned (textvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-textvqa-224-jax | |
| paligemma-3b-ft-textvqa-448 | fine-tuned (textvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-textvqa-448-jax | |
| paligemma-3b-ft-textvqa-896 | fine-tuned (textvqa, 896px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-textvqa-896-jax | |
| paligemma-3b-ft-vizwizvqa-224 | fine-tuned (vizwizvqa, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-vizwizvqa-224-jax | |
| paligemma-3b-ft-vizwizvqa-448 | fine-tuned (vizwizvqa, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-vizwizvqa-448-jax | |
| paligemma-3b-ft-vqav2-224 | fine-tuned (vqav2, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-vqav2-224-jax | |
| paligemma-3b-ft-vqav2-448 | fine-tuned (vqav2, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-vqav2-448-jax | |
| paligemma-3b-ft-widgetcap-224 | fine-tuned (widgetcap, 224px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-widgetcap-224-jax | |
| paligemma-3b-ft-widgetcap-448 | fine-tuned (widgetcap, 448px) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-ft-widgetcap-448-jax | |
| paligemma-3b-mix-224 | instruct (mix) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-mix-224-jax · paligemma-3b-mix-224-keras | |
| paligemma-3b-mix-448 | instruct (mix) | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-mix-448-jax · paligemma-3b-mix-448-keras | |
| paligemma-3b-pt-224 | base | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-pt-224-jax · paligemma-3b-pt-224-keras | |
| paligemma-3b-pt-448 | base | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-pt-448-jax · paligemma-3b-pt-448-keras | |
| paligemma-3b-pt-896 | base | 3B | 2024-05-14 | gemma (Gemma Terms of Use) | Download (gated) · paligemma-3b-pt-896-jax · paligemma-3b-pt-896-keras |
RecurrentGemma from 2024-04-09
Griffin-architecture models (recurrence plus local attention) released alongside Gemma 1; gemma.ai's family tree doesn't place them under a generation. Gated behind the Gemma Terms of Use.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| recurrentgemma-2b | base | 2B | 2024-04-09 | gemma (Gemma Terms of Use) | Download (gated) · recurrentgemma-2b-flax · recurrentgemma-2b-sfp-cpp | |
| recurrentgemma-2b-it | instruct | 2B | 2024-04-09 | gemma (Gemma Terms of Use) | Download (gated) · recurrentgemma-2b-it-flax · recurrentgemma-2b-it-sfp-cpp | |
| recurrentgemma-9b | base | 9B | 2024-06-11 | gemma (Gemma Terms of Use) | Download (gated) | |
| recurrentgemma-9b-it | instruct | 9B | 2024-06-11 | gemma (Gemma Terms of Use) | Download (gated) |
VaultGemma from 2025-09-13
1B model trained with differential privacy; gemma.ai's family tree doesn't place it under a generation and its card is behind the gate, so we couldn't confirm which generation it derives from. Gated behind the Gemma Terms of Use.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| vaultgemma-1b | base | 1B | 2025-09-13 | gemma (Gemma Terms of Use) | Download (gated) |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: DiffusionGemma 26B A4B IT 5, Gemma 4 12B IT 5, Gemma 4 26B A4B IT 5, Gemma 4 31B IT 5, Gemma 4 E2B IT 5, Gemma 4 E4B IT 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Hugging Face (SmolLM)
Publishes from huggingface.co/HuggingFaceTB. Checked 2026-10-21.
SmolLM3 from 2025-07-08
Only one size (3B). Dates are our listing's firstReleased for SmolLM3-3B and otherwise the repository's createdAt from the Hub API (SmolLM3-3B-Base was created 2025-06-19, likely private before the 8 July release). Nothing is gated. Licence is apache-2.0 throughout.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| SmolLM3-3B SmolLM3-3B is Hugging Face's open-weight dense 3B instruct model with switchable reasoning, tool calling and a 65,536-token native context, published in BF16 safetensors under Apache 2.0 with its training data and recipes. | instruct (reasoning on/off) | 3B | 2025-07-08 | apache-2.0 | HuggingFaceTB/SmolLM3-3B-Base | Our listing (C) · Download · SmolLM3-3B-ONNX |
| SmolLM3-3B-Base | base | 3B | 2025-06-19 | apache-2.0 | Download | |
| SmolLM3-3B-checkpoints | intermediate training checkpoints (research) | 3B | 2025-07-20 | apache-2.0 | HuggingFaceTB/SmolLM3-3B-Base | Download |
| SmolLM3-3B-GSM8K-SFT | maths fine-tune (GSM8K SFT) | 3B | 2026-04-03 | apache-2.0 | HuggingFaceTB/SmolLM3-3B-Base | Download |
Older versions (2 families, 20 models)
SmolLM2 from 2024-10-31
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| SmolLM2-1.7B-Instruct | instruct | 1.7B | 2024-10-31 | apache-2.0 | HuggingFaceTB/SmolLM2-1.7B | Download · SmolLM2-1.7B-Instruct-GGUF · SmolLM2-1.7B-Instruct-Q8-mlx |
| SmolLM2-1.7B | base | 1.7B | 2024-10-30 | apache-2.0 | Download | |
| SmolLM2-1.7B-Instruct-16k | instruct, 16K context | 1.7B | 2025-02-21 | apache-2.0 | HuggingFaceTB/SmolLM2-1.7B-Instruct | Download |
| SmolLM2-1.7B-sft-only (Front matter names HuggingFaceTB/SmolLM2-1.7B-8k, which is not a public repository.) | SFT-only checkpoint | 1.7B | 2024-10-30 | not stated on card or front matter | HuggingFaceTB/SmolLM2-1.7B-8k | Download |
| SmolLM2-360M-Instruct | instruct | 360M | 2024-10-31 | apache-2.0 | HuggingFaceTB/SmolLM2-360M | Download · SmolLM2-360M-Instruct-GGUF · SmolLM2-360M-Instruct-Q8-mlx |
| SmolLM2-360M | base | 360M | 2024-10-31 | apache-2.0 | Download | |
| SmolLM2-135M-Instruct | instruct | 135M | 2024-10-31 | apache-2.0 | HuggingFaceTB/SmolLM2-135M | Download · SmolLM2-135M-Instruct-Q8-mlx |
| SmolLM2-135M | base | 135M | 2024-10-31 | apache-2.0 | Download | |
| smollm2-135M-SFT-Only (Front matter names HuggingFaceTB/135M-lc-100k-rope-12B, which is not a public repository.) | SFT-only checkpoint | 135M | 2024-10-31 | not stated on card or front matter | HuggingFaceTB/135M-lc-100k-rope-12B | Download |
| SmolLM2-1.7B-intermediate-checkpoints | intermediate training checkpoints (research) | 1.7B | 2025-02-26 | apache-2.0 | Download | |
| SmolLM2-360M-intermediate-checkpoints | intermediate training checkpoints (research) | 360M | 2025-02-26 | apache-2.0 | Download | |
| SmolLM2-135M-intermediate-checkpoints | intermediate training checkpoints (research) | 135M | 2025-02-26 | apache-2.0 | Download | |
| SmolLM2-nanotron-ckpt | training checkpoints in nanotron format | 135M, 360M, 1.7B | 2024-11-30 | apache-2.0 | Download |
SmolLM from 2024-07-14
The Instruct repositories were updated in place to v0.2 in August 2024; the add-basics repositories hold the v0.2 fine-tunes that the Q8 GGUF and MLC builds were made from.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| SmolLM-1.7B-Instruct | instruct | 1.7B | 2024-07-15 | apache-2.0 | HuggingFaceTB/SmolLM-1.7B | Download |
| SmolLM-1.7B | base | 1.7B | 2024-07-14 | apache-2.0 | Download | |
| SmolLM-360M-Instruct | instruct | 360M | 2024-07-15 | apache-2.0 | HuggingFaceTB/SmolLM-360M | Download · SmolLM-360M-Instruct-ONNX-fp16 |
| smollm-360M-instruct-add-basics | instruct (v0.2 fine-tune) | 360M | 2024-08-13 | apache-2.0 | HuggingFaceTB/SmolLM-360M | Download · smollm-360M-instruct-v0.2-Q8_0-GGUF · smollm-360M-instruct-add-basics-q0f16-MLC |
| SmolLM-360M | base | 360M | 2024-07-14 | apache-2.0 | Download | |
| SmolLM-135M-Instruct | instruct | 135M | 2024-07-15 | apache-2.0 | HuggingFaceTB/SmolLM-135M | Download · smollm-135M-instruct-v0.2-Q8_0-GGUF · smollm-135M-instruct-add-basics-q0f16-MLC |
| SmolLM-135M | base | 135M | 2024-07-14 | apache-2.0 | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: SmolLM3-3B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
IBM (Granite)
Publishes from huggingface.co/ibm-granite. Checked 2026-10-21.
Granite 4 (4.0, 4.1, 4.2) from 2025-04-30
IBM's current language models: 4.0 (tiny preview in April 2025, then micro, tiny, small and the hybrid Mamba "h" sizes from 350M to 32B, September and October 2025), 4.1 (3B, 8B, 30B with bases, April 2026) and 4.2 (3B, 8B, 30B, 7 August 2026, each a fine-tune of the 4.1 base, with a thinking mode). IBM's own GGUF, FP8, MXFP4, NVFP4 and MLX repositories are builds. Granite 4.1 3B (granite-4-1-3b) was in the first draft of the second batch and was replaced by Granite 4.2 3B, so it has no listing. Vision and speech models built on Granite 4 are in their own families below. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-4.2-3b Granite 4.2 3B is IBM's open-weight dense reasoning model of 3.7B parameters, post-trained from Granite 4.1 3B Base, with switchable thinking, tool calling and a 131,072-token context, published in BF16, FP8, FP4, GGUF and MLX under Apache 2.0. (Released 25 August 2026 by the card and IBM's blog. The repository was created on 7 August 2026) | instruct (with a thinking mode) | 3B (Hub safetensors count 3.66B) | 2026-08-25 | apache-2.0 | Our listing (C) · Download · granite-4.2-3b-GGUF · granite-4.2-3b-fp8 · granite-4.2-3b-mxfp4 · granite-4.2-3b-nvfp4 · granite-4.2-3b-q4-mlx · granite-4.2-3b-q8-mlx · granite-4.2-3b-bf16-mlx | |
| granite-4.2-8b | instruct (with a thinking mode) | 8B (Hub safetensors count 8.79B) | 2026-08-07 | apache-2.0 | Download · granite-4.2-8b-GGUF · granite-4.2-8b-fp8 · granite-4.2-8b-mxfp4 · granite-4.2-8b-nvfp4 · granite-4.2-8b-q4-mlx · granite-4.2-8b-q8-mlx · granite-4.2-8b-bf16-mlx | |
| granite-4.2-30b | instruct (with a thinking mode) | 30B (Hub safetensors count 29.28B) | 2026-08-07 | apache-2.0 | Download · granite-4.2-30b-GGUF · granite-4.2-30b-fp8 · granite-4.2-30b-mxfp4 · granite-4.2-30b-nvfp4 · granite-4.2-30b-q4-mlx · granite-4.2-30b-q8-mlx · granite-4.2-30b-bf16-mlx | |
| granite-4.1-8b-base | base | 8B (Hub safetensors count 8.38B) | 2026-04-06 | apache-2.0 | Download | |
| granite-4.1-8b | instruct | 8B (Hub safetensors count 8.79B) | 2026-04-06 | apache-2.0 | Download · granite-4.1-8b-GGUF · granite-4.1-8b-fp8 | |
| granite-4.1-3b-base | base | 3B (Hub safetensors count 3.40B) | 2026-04-06 | apache-2.0 | Download | |
| granite-4.1-3b | instruct | 3B (Hub safetensors count 3.40B) | 2026-04-06 | apache-2.0 | Download · granite-4.1-3b-GGUF · granite-4.1-3b-fp8 | |
| granite-4.1-30b-base | base | 30B (Hub safetensors count 28.87B) | 2026-04-06 | apache-2.0 | Download | |
| granite-4.1-30b | instruct | 30B (Hub safetensors count 28.87B) | 2026-04-06 | apache-2.0 | Download · granite-4.1-30b-GGUF · granite-4.1-30b-fp8 | |
| granite-4.0-h-350m-base | base | 350M (Hub safetensors count 340M) | 2025-10-07 | apache-2.0 | Download · granite-4.0-h-350m-base-GGUF | |
| granite-4.0-h-350m | instruct | 350M (Hub safetensors count 340M) | 2025-10-07 | apache-2.0 | Download · granite-4.0-h-350m-GGUF | |
| granite-4.0-h-1b-base | base | 1B (Hub safetensors count 1.46B) | 2025-10-07 | apache-2.0 | Download · granite-4.0-h-1b-base-GGUF | |
| granite-4.0-h-1b | instruct | 1B (Hub safetensors count 1.46B) | 2025-10-07 | apache-2.0 | Download · granite-4.0-h-1b-GGUF | |
| granite-4.0-350m-base | base | 350M (Hub safetensors count 352M) | 2025-10-07 | apache-2.0 | Download · granite-4.0-350m-base-GGUF | |
| granite-4.0-350m | instruct | 350M (Hub safetensors count 352M) | 2025-10-07 | apache-2.0 | Download · granite-4.0-350m-GGUF | |
| granite-4.0-1b-base | base | 1B (Hub safetensors count 1.63B) | 2025-10-07 | apache-2.0 | Download · granite-4.0-1b-base-GGUF | |
| granite-4.0-1b | instruct | 1B (Hub safetensors count 1.63B) | 2025-10-07 | apache-2.0 | Download · granite-4.0-1b-GGUF | |
| granite-4.0-micro-base | base | 3.40B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-micro-base-GGUF | |
| granite-4.0-micro | instruct | 3.40B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-micro-GGUF | |
| granite-4.0-h-tiny-base | base | 6.94B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-tiny-base-GGUF | |
| granite-4.0-h-tiny | instruct | 6.94B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-tiny-GGUF | |
| granite-4.0-h-small-base | base | 32.21B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-small-base-GGUF | |
| granite-4.0-h-small | instruct | 32.21B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-small-GGUF · granite-4.0-h-small-FP8 | |
| granite-4.0-h-micro-base | base | 3.19B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-micro-base-GGUF | |
| granite-4.0-h-micro | instruct | 3.19B (Hub safetensors count; the name states none) | 2025-09-16 | apache-2.0 | Download · granite-4.0-h-micro-GGUF | |
| granite-4.0-tiny-preview | instruct | 6.67B (Hub safetensors count; the name states none) | 2025-04-30 | apache-2.0 | Download · granite-4.0-tiny-preview-GGUF | |
| granite-4.0-tiny-base-preview | base | 6.67B (Hub safetensors count; the name states none) | 2025-04-30 | apache-2.0 | Download · granite-4.0-tiny-base-preview-GGUF |
Granite Embedding R2 from 2025-07-17
Second-release Granite Embedding (ModernBERT-based, English R2 in July 2025 and multilingual R2 in April 2026) and the R2 reranker. R1 is listed under older.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-embedding-small-english-r2 Granite Embedding Small English R2 is IBM's open-weight English text embedding model, a 47M-parameter ModernBERT bi-encoder that returns 384-dimension vectors for inputs of up to 8,192 tokens, trained on permissively licensed and IBM data, under Apache 2.0. | embedding | 48M (Hub safetensors count; the name states none) | 2025-07-17 | apache-2.0 | Our listing (E) · Download | |
| granite-embedding-97m-multilingual-r2 | embedding | 97M (Hub safetensors count 97M) | 2026-04-20 | apache-2.0 | Download | |
| granite-embedding-311m-multilingual-r2 | embedding | 311M (Hub safetensors count 312M) | 2026-04-20 | apache-2.0 | Download | |
| granite-embedding-reranker-english-r2 | reranker | 150M (Hub safetensors count; the name states none) | 2025-08-04 | apache-2.0 | Download | |
| granite-embedding-english-r2 | embedding | 149M (Hub safetensors count; the name states none) | 2025-07-17 | apache-2.0 | Download |
Granite Guardian from 2024-09-05
IBM's judge models, from the HAP classifiers (38M and 125M, September 2024) and Guardian 3.0 (October 2024) to Guardian 4.1 8B (April 2026), which our granite-guardian listing grades. LoRA adapters for Guardian 3.2 are left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-guardian-4.1-8b Granite Guardian is IBM's family of open-weight judge models. The current 8-billion-parameter release answers yes or no on whether a prompt, response, retrieved context or function call meets a built-in or custom criterion, and the owner runs it. | safety classifier (judge) | 8B (Hub safetensors count 8.38B) | 2026-04-16 | apache-2.0 | ibm-granite/granite-4.1-8b | Our listing (C) · Download · granite-guardian-4.1-8b-GGUF |
| granite-guardian-4.0-3b-toxicity-ja | safety classifier (Japanese toxicity) | 3B (Hub safetensors count 3.40B) | 2026-04-06 | apache-2.0 | ibm-granite/granite-4.0-micro | Download |
| granite-guardian-3.2-8b-factuality-detection | safety classifier (factuality detection) | 8B (Hub safetensors count 8.17B) | 2025-11-10 | apache-2.0 | ibm-granite/granite-3.2-8b-instruct | Download |
| granite-guardian-3.3-8b | safety classifier (judge) | 8B (Hub safetensors count 8.17B) | 2025-06-03 | apache-2.0 | Download · granite-guardian-3.3-8b-GGUF | |
| granite-guardian-3.2-3b-a800m | safety classifier (judge) | 3B total, 800M active (MoE) (Hub safetensors count 3.30B) | 2025-02-03 | apache-2.0 | Download | |
| granite-guardian-3.2-5b | safety classifier (judge) | 5B (Hub safetensors count 5.78B) | 2025-01-23 | apache-2.0 | Download | |
| granite-guardian-3.1-8b | safety classifier (judge) | 8B (Hub safetensors count 8.17B) | 2024-12-17 | apache-2.0 | Download | |
| granite-guardian-3.1-2b | safety classifier (judge) | 2B (Hub safetensors count 2.53B) | 2024-12-17 | apache-2.0 | Download | |
| granite-guardian-3.0-8b | safety classifier (judge) | 8B (Hub safetensors count 8.17B) | 2024-10-15 | apache-2.0 | Download | |
| granite-guardian-3.0-2b | safety classifier (judge) | 2B (Hub safetensors count 2.53B) | 2024-10-15 | apache-2.0 | Download | |
| granite-guardian-hap-38m | classifier (hate, abuse and profanity) | 38M (Hub safetensors count 38M) | 2024-09-05 | apache-2.0 | Download | |
| granite-guardian-hap-125m | classifier (hate, abuse and profanity) | 125M (Hub safetensors count 125M) | 2024-09-05 | apache-2.0 | Download |
Granite Switch from 2026-05-01
Granite 4.1 checkpoints composed with decision and guardian aLoRA adapters and a Granite Speech encoder, served as one endpoint; previews in May 2026 and the 8B and 30B releases on 14 October 2026. Our granite-switch listing covers the 8B and 30B. The card declares no base_model; the base is taken from our listing.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-switch-4.1-8b Granite Switch 4.1 is IBM's open-weight 8B and 30B Granite 4.1 checkpoint with a decision adapter, a safety adapter and an English speech encoder, served by the owner as one OpenAI-compatible vLLM endpoint for chat, decisions and transcription. | instruct with decision and guardian adapters and a speech encoder in one checkpoint | 8B (Hub safetensors count 8.78B) | 2026-10-14 | apache-2.0 | ibm-granite/granite-4.1-8b | Our listing (C) · Download |
| granite-switch-4.1-30b Granite Switch 4.1 is IBM's open-weight 8B and 30B Granite 4.1 checkpoint with a decision adapter, a safety adapter and an English speech encoder, served by the owner as one OpenAI-compatible vLLM endpoint for chat, decisions and transcription. | instruct with decision and guardian adapters and a speech encoder in one checkpoint | 30B (Hub safetensors count 29.99B) | 2026-10-14 | apache-2.0 | ibm-granite/granite-4.1-30b | Our listing (C) · Download |
| granite-switch-4.1-8b-preview | instruct with adapters in one checkpoint (preview) | 8B (Hub safetensors count 9.57B) | 2026-05-01 | apache-2.0 | Download | |
| granite-switch-4.1-3b-preview | instruct with adapters in one checkpoint (preview) | 3B (Hub safetensors count 4.15B) | 2026-05-01 | apache-2.0 | Download | |
| granite-switch-4.1-30b-preview | instruct with adapters in one checkpoint (preview) | 30B (Hub safetensors count 32.24B) | 2026-05-01 | apache-2.0 | Download |
Older versions (7 families, 64 models)
Granite Swash from 2026-07-01
Two July 2026 models whose cards were not read; edition unknown.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-swash-3b-a600m | not stated (card not read) | 3B total, 600M active (MoE) (Hub safetensors count 3.02B) | 2026-07-01 | apache-2.0 | Download | |
| granite-swash-2b | not stated (card not read) | 2B (Hub safetensors count 2.14B) | 2026-07-01 | apache-2.0 | Download |
Granite Speech from 2025-03-26
Speech recognisers on Granite 3.2, 3.3 and 4.0 language models, plus Granite Speech 5.0 TurboCTC (470M; the -nc edition is cc-by-nc-sa-4.0).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-speech-5.0-470m-turboctc-nc | speech-to-text | 470M (Hub safetensors count 473M) | 2026-08-17 | cc-by-nc-sa-4.0 | Download | |
| granite-speech-5.0-470m-turboctc | speech-to-text | 470M (Hub safetensors count 473M) | 2026-08-04 | apache-2.0 | Download | |
| granite-speech-4.1-2b-plus | speech-to-text | 2B (Hub safetensors count 2.11B) | 2026-04-16 | apache-2.0 | ibm-granite/granite-4.0-1b-base | Download · granite-speech-4.1-2b-plus-GGUF |
| granite-speech-4.1-2b | speech-to-text | 2B (Hub safetensors count 2.31B) | 2026-04-16 | apache-2.0 | ibm-granite/granite-4.0-1b-base | Download · granite-speech-4.1-2b-GGUF |
| granite-speech-4.1-2b-nar | speech-to-text (non-autoregressive) | 2B (Hub safetensors count 2.25B) | 2026-03-10 | apache-2.0 | ibm-granite/granite-4.0-1b-base | Download |
| granite-4.0-1b-speech | speech-to-text | 1B (Hub safetensors count 2.31B) | 2026-02-27 | apache-2.0 | ibm-granite/granite-4.0-1b-base | Download · granite-4.0-1b-speech-GGUF |
| granite-speech-3.3-2b | speech-to-text | 2B (Hub safetensors count 3.01B) | 2025-04-28 | apache-2.0 | ibm-granite/granite-3.3-2b-instruct | Download |
| granite-speech-3.3-8b | speech-to-text | 8B (Hub safetensors count 8.65B) | 2025-04-14 | apache-2.0 | ibm-granite/granite-3.3-8b-instruct | Download |
| granite-speech-3.2-8b | speech-to-text | 8B (Hub safetensors count 8.48B) | 2025-03-26 | apache-2.0 | ibm-granite/granite-3.2-8b-instruct | Download |
Granite Vision from 2025-01-27
Vision-language models on Granite 3.1 and 4.x, including a visual-document embedding model and a chart-to-CSV preview.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-vision-4.1-4b | instruct (vision-language) | 4B (Hub safetensors count 4.00B) | 2026-04-16 | apache-2.0 | Download · granite-vision-4.1-4b-GGUF | |
| granite-4.0-3b-vision | instruct (vision-language) | 3B (Hub safetensors count 4.00B) | 2026-03-03 | apache-2.0 | Download | |
| granite-vision-3.3-2b-chart2csv-preview | instruct (vision-language) | 2B (Hub safetensors count 2.98B) | 2026-01-29 | apache-2.0 | Download · granite-vision-3.3-2b-chart2csv-preview-GGUF | |
| granite-vision-3.3-2b-embedding | embedding (visual documents) | 2B (Hub safetensors count 2.98B) | 2025-06-03 | apache-2.0 | Download | |
| granite-vision-3.3-2b | instruct (vision-language) | 2B (Hub safetensors count 2.98B) | 2025-06-03 | apache-2.0 | Download · granite-vision-3.3-2b-GGUF | |
| granite-vision-3.2-2b | instruct (vision-language) | 2B (Hub safetensors count 2.98B) | 2025-02-17 | apache-2.0 | ibm-granite/granite-3.1-2b-instruct | Download |
| granite-vision-3.1-2b-preview | instruct (vision-language) | 2B (Hub safetensors count 2.98B) | 2025-01-27 | apache-2.0 | ibm-granite/granite-3.1-2b-instruct | Download |
Granite Embedding (R1) from 2024-12-04
First-release Granite Embedding (December 2024), including a sparse model.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-embedding-30m-sparse | embedding | 30M (Hub safetensors count 30M) | 2025-02-17 | apache-2.0 | Download | |
| granite-embedding-30m-english | embedding | 30M (Hub safetensors count 30M) | 2024-12-04 | apache-2.0 | Download | |
| granite-embedding-278m-multilingual | embedding | 278M (Hub safetensors count 278M) | 2024-12-04 | apache-2.0 | Download | |
| granite-embedding-125m-english | embedding | 125M (Hub safetensors count 125M) | 2024-12-04 | apache-2.0 | Download | |
| granite-embedding-107m-multilingual | embedding | 107M (Hub safetensors count 107M) | 2024-12-04 | apache-2.0 | Download |
Granite 3 (3.0 to 3.3) from 2024-10-02
Dense and MoE language models from October 2024 to April 2025, plus a 2026 process reward model on 3.3 8B. LoRA and aLoRA adapters, intrinsics libraries and speculative-decoding accelerators are left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-3.3-8b-math-prm-v2 | reward model (math process reward) | 8B (Hub safetensors count 8.17B) | 2026-01-07 | apache-2.0 | Download | |
| granite-3.3-8b-instruct | instruct | 8B (Hub safetensors count 8.17B) | 2025-04-09 | apache-2.0 | Download · granite-3.3-8b-instruct-GGUF · granite-3.3-8b-instruct-FP8 | |
| granite-3.3-8b-base | base | 8B (Hub safetensors count 8.17B) | 2025-04-09 | apache-2.0 | Download · granite-3.3-8b-base-GGUF | |
| granite-3.3-2b-instruct | instruct | 2B (Hub safetensors count 2.53B) | 2025-04-09 | apache-2.0 | Download · granite-3.3-2b-instruct-GGUF | |
| granite-3.3-2b-base | base | 2B (Hub safetensors count 2.53B) | 2025-04-09 | apache-2.0 | Download · granite-3.3-2b-base-GGUF | |
| granite-3.2-8b-instruct | instruct | 8B (Hub safetensors count 8.17B) | 2025-02-17 | apache-2.0 | Download | |
| granite-3.2-2b-instruct | instruct | 2B (Hub safetensors count 2.53B) | 2025-02-17 | apache-2.0 | Download | |
| granite-3.2-8b-instruct-preview | instruct | 8B (Hub safetensors count 8.17B) | 2025-02-07 | apache-2.0 | Download | |
| granite-3.1-8b-instruct | instruct | 8B (Hub safetensors count 8.17B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-8b-base | base | 8B (Hub safetensors count 8.17B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-3b-a800m-instruct | instruct | 3B total, 800M active (MoE) (Hub safetensors count 3.30B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-3b-a800m-base | base | 3B total, 800M active (MoE) (Hub safetensors count 3.30B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-2b-instruct | instruct | 2B (Hub safetensors count 2.53B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-2b-base | base | 2B (Hub safetensors count 2.53B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-1b-a400m-instruct | instruct | 1B total, 400M active (MoE) (Hub safetensors count 1.33B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.1-1b-a400m-base | base | 1B total, 400M active (MoE) (Hub safetensors count 1.33B) | 2024-12-06 | apache-2.0 | Download | |
| granite-3.0-3b-a800m-instruct | instruct | 3B total, 800M active (MoE) (Hub safetensors count 3.37B) | 2024-10-03 | apache-2.0 | Download | |
| granite-3.0-3b-a800m-base | base | 3B total, 800M active (MoE) (Hub safetensors count 3.37B) | 2024-10-03 | apache-2.0 | Download | |
| granite-3.0-1b-a400m-instruct | instruct | 1B total, 400M active (MoE) (Hub safetensors count 1.33B) | 2024-10-03 | apache-2.0 | Download | |
| granite-3.0-1b-a400m-base | base | 1B total, 400M active (MoE) (Hub safetensors count 1.38B) | 2024-10-03 | apache-2.0 | Download | |
| granite-3.0-8b-instruct | instruct | 8B (Hub safetensors count 8.17B) | 2024-10-02 | apache-2.0 | Download | |
| granite-3.0-8b-base | base | 8B (Hub safetensors count 8.17B) | 2024-10-02 | apache-2.0 | Download | |
| granite-3.0-2b-instruct | instruct | 2B (Hub safetensors count 2.63B) | 2024-10-02 | apache-2.0 | Download | |
| granite-3.0-2b-base | base | 2B (Hub safetensors count 2.53B) | 2024-10-02 | apache-2.0 | Download |
Granite Code from 2024-04-21
Code models from April to July 2024 (3B to 34B, base and instruct, 4K to 128K context) and a 20B function-calling model.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| granite-8b-code-instruct-128k | instruct | 8B (Hub safetensors count 8.05B) | 2024-07-12 | apache-2.0 | Download | |
| granite-3b-code-instruct-128k | instruct | 3B (Hub safetensors count 3.48B) | 2024-07-12 | apache-2.0 | Download | |
| granite-20b-functioncalling | instruct (function calling) | 20B (Hub safetensors count 20.07B) | 2024-07-09 | apache-2.0 | Download | |
| granite-20b-code-instruct-r1.1 | instruct | 20B (Hub safetensors count 20.07B) | 2024-07-09 | apache-2.0 | Download | |
| granite-20b-code-base-r1.1 | base | 20B (Hub safetensors count 20.07B) | 2024-07-09 | apache-2.0 | Download | |
| granite-8b-code-base-128k | base | 8B (Hub safetensors count 8.05B) | 2024-06-29 | apache-2.0 | Download | |
| granite-3b-code-base-128k | base | 3B (Hub safetensors count 3.48B) | 2024-06-29 | apache-2.0 | Download | |
| granite-8b-code-instruct-4k | instruct | 8B (Hub safetensors count 8.05B) | 2024-04-26 | apache-2.0 | Download · granite-8b-code-instruct-4k-GGUF | |
| granite-3b-code-instruct-2k | instruct | 3B (Hub safetensors count 3.48B) | 2024-04-26 | apache-2.0 | Download · granite-3b-code-instruct-2k-GGUF | |
| granite-34b-code-instruct-8k | instruct | 34B (Hub safetensors count 33.72B) | 2024-04-26 | apache-2.0 | Download · granite-34b-code-instruct-8k-GGUF | |
| granite-20b-code-instruct-8k | instruct | 20B (Hub safetensors count 20.07B) | 2024-04-26 | apache-2.0 | Download · granite-20b-code-instruct-8k-GGUF | |
| granite-3b-code-base-2k | base | 3B (Hub safetensors count 3.48B) | 2024-04-23 | apache-2.0 | Download · granite-3b-code-base-2k-GGUF | |
| granite-8b-code-base-4k | base | 8B (Hub safetensors count 8.05B) | 2024-04-21 | apache-2.0 | Download · granite-8b-code-base-4k-GGUF | |
| granite-34b-code-base-8k | base | 34B (Hub safetensors count 33.72B) | 2024-04-21 | apache-2.0 | Download · granite-34b-code-base-8k-GGUF | |
| granite-20b-code-base-8k | base | 20B (Hub safetensors count 20.07B) | 2024-04-21 | apache-2.0 | Download · granite-20b-code-base-8k-GGUF |
Granite 7B (2024) from 2024-04-19
The first Granite language models published on the Hub.
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Granite 4.2 3B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
inclusionAI (Ant Group)
Publishes from huggingface.co/inclusionAI. Checked 2026-10-21.
Ming 3.0 from 2026-10-19
inclusionAI's current Ming speech line: two text-to-speech checkpoints, two audio-language models and one ASR model, all MIT and ungated. The Hugging Face repositories were created between 3 September and 14 October 2026 (createdAt) but our listings date the public release to 19 October 2026, which is used here. No Ming-3.0 omni (vision) checkpoint is published yet. inclusionAI publishes no quantised builds of these in its own organisation; ModelScope mirrors are linked from the cards.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-3.0-tts-0.5B Ming-3.0-tts-0.5B is inclusionAI's open-weight text-to-speech model on a Qwen2.5-0.5B backbone, with voice cloning, voice design, built-in voices and inline paralinguistic tags in seven languages, published under MIT and run with inclusionAI's Python code. | text-to-speech | 0.5B (dense) | 2026-10-19 | mit | Qwen2.5-0.5B | Our listing (C) · Download |
| Ming-3.0-tts-tiny Ming-3.0-tts-tiny is inclusionAI's open-weight text-to-speech model, a Ling V3 mixture of experts with 7.9B backbone parameters and about 1.3B active, for voice cloning, voice design and controlled speech in seven languages, under MIT. | text-to-speech | 7.9B total, about 1.3B active (backbone) | 2026-10-19 | mit | Ling V3 (7.9B-A1.3B backbone, per card) | Our listing (D) · Download |
| Ming-3.0-audio-flash Ming-3.0-audio-flash is inclusionAI's open-weight audio-language model for Chinese and English. It takes audio with text instructions and returns transcripts, answers about the audio, or spoken replies, 126B parameters with 6.5B active, under MIT. | audio-language (speech recognition, audio understanding, optional speech output) | 126B total, 6.5B active | 2026-10-19 | mit | Our listing (E) · Download | |
| Ming-3.0-audio-tiny Ming-3.0-audio-tiny is inclusionAI's open-weight audio-language model for Chinese and English speech recognition, audio understanding and spoken replies, 9.3B parameters with 2.7B active, published on Hugging Face under MIT. | audio-language (speech recognition, audio understanding, optional speech output) | 9.3B total, 2.7B active | 2026-10-19 | mit | Our listing (D) · Download | |
| Ming-3.0-asr-tiny Ming-3.0-asr-tiny is inclusionAI's open-weight speech recognition model for Chinese, English and Chinese dialects, an audio encoder on the Ling-tiny-v3 mixture of experts with 8.9B parameters counted from the files, that takes hotwords or dialogue context as hints, under MIT. | speech recognition | not stated on the card (Hub metadata counts 8,865,800,992 in the files) | 2026-10-19 | mit | Ling-tiny-v3 (language backbone, per card) | Our listing (D) · Download |
Ling 3 from 2026-08-02
inclusionAI's current language-model family (Ling-3.0 and Ling-3.1, minor versions of one major). No listing with us; included as its own family at the caller's request, and because the Ming 3.0 tiny models name a Ling V3 / Ling-tiny-v3 backbone. Ling 3 cards describe the models as native hybrid reasoning models; there is no separate Ring-3 reasoning release. Also published: SingProbe guardrail probes and a DSpark speculative-decoding draft, which are add-ons to Ling-3.0-flash/tiny rather than standalone models.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ling-3.0-flash | instruct (hybrid reasoning) | 124B total, 5.1B active | 2026-08-02 | mit | Download · Ling-3.0-flash-fp8 · Ling-3.0-flash-int4 · Ling-3.0-flash-fp4 · Ling-3.0-flash-GGUF | |
| Ling-3.0-flash-base | base | 124B total, 5.1B active | 2026-08-11 | mit | Download | |
| Ling-3.0-flash-base-30T | base (30T-token checkpoint) | 124B total, 5.1B active | 2026-08-11 | mit | Download | |
| Ling-3.0-flash-base-midtrain | base (mid-training checkpoint) | 124B total, 5.1B active | 2026-08-11 | mit | Download | |
| Ling-3.0-flash-Fin | instruct (finance) | 124B total, 5.1B active | 2026-09-03 | mit | inclusionAI/Ling-3.0-flash | Download · Ling-3.0-flash-Fin-fp8 · Ling-3.0-flash-Fin-int4 · Ling-3.0-flash-Fin-fp4 |
| Ling-3.0-flash-VL | vision-language instruct | 124B total, 5.5B active | 2026-09-04 | mit | inclusionAI/Ling-3.0-flash | Download · Ling-3.0-flash-VL-fp8 · Ling-3.0-flash-VL-fp4 · Ling-3.0-flash-VL-int4 |
| Ling-3.0-flash-dspark | speculative-decoding draft model | 1.36B (draft) | 2026-08-09 | mit | inclusionAI/Ling-3.0-flash (target model) | Download |
| Ling-3.0-flash-singprobe | guardrail probe on the base model's hidden states | 5.18M (probe) | 2026-09-01 | mit | inclusionAI/Ling-3.0-flash | Download |
| Ling-3.0-tiny Ling-3.0-tiny is inclusionAI's open-weight hybrid reasoning mixture of experts with 7.9B parameters and 1.3B active, mixing Kimi Delta Attention and MLA layers 3 to 1, with switchable thinking and a 131,072-token context, in BF16, FP8, INT4 and GGUF under MIT. | instruct (hybrid reasoning) | 7.9B total, 1.3B active | 2026-08-10 | mit | Our listing (C) · Download · Ling-3.0-tiny-fp8 · Ling-3.0-tiny-int4 · Ling-3.0-tiny-GGUF | |
| Ling-3.0-tiny-base | base | 7.9B total, 1.3B active | 2026-08-11 | mit | Download | |
| Ling-3.0-tiny-base-30T | base (30T-token checkpoint) | 7.9B total, 1.3B active | 2026-08-11 | mit | Download | |
| Ling-3.0-tiny-base-midtrain | base (mid-training checkpoint) | 7.9B total, 1.3B active | 2026-08-11 | mit | Download | |
| Ling-3.0-tiny-singprobe | guardrail probe on the base model's hidden states | 3.22M (probe) | 2026-09-01 | mit | inclusionAI/Ling-3.0-tiny | Download |
| Ling-3.1-flash | instruct (hybrid reasoning) | 560B total, 25B active | 2026-10-13 | mit | Ling-3.0-flash (card: 'builds upon Ling-3.0-flash') | Download |
Older versions (10 families, 54 models)
Ming-Image 0.1 from 2026-09-17
Parallel Ming image-generation line released a month before Ming 3.0, not a predecessor of it.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-Image-0.1-Design | text-to-image | 6B | 2026-09-17 | mit | Download | |
| Ming-Image-0.1-Design-Layer | image-to-image (layered design) | not stated on the card | 2026-09-17 | mit | Download |
Ming-flash-omni 2.0 and Ming-omni-tts from 2026-02-11
Released together in February 2026. Only Ming-flash-omni carries the 2.0 number; grouping Ming-omni-tts with it is ours.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-flash-omni-2.0 | omni (any-to-any) | 100B total, 6B active | 2026-02-11 | mit | Ling-2.0 MoE architecture (card names the architecture, not a checkpoint) | Download |
| Ming-omni-tts-16.8B-A3B | text-to-speech (speech, sound and music) | 16.8B total, 3B active (from the repository name) | 2026-02-11 | apache-2.0 | Download | |
| Ming-omni-tts-0.5B | text-to-speech (speech, sound and music) | 0.5B (from the repository name) | 2026-02-11 | apache-2.0 | Download | |
| Ming-omni-tts-tokenizer-12Hz | audio tokenizer | not stated on the card | 2026-02-11 | apache-2.0 | Download |
Ming-flash-omni Preview from 2025-10-27
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-flash-omni-Preview | omni (any-to-any), preview | 100B total, 6B active | 2025-10-27 | mit | Ling-Flash-2.0 (sparser MoE variant, per card) | Download |
Ming-VideoMAR from 2025-10-17
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-VideoMAR (Repository holds .pth checkpoints and no licence field.) | image-to-video | not stated on the card | 2025-10-17 | Download |
Ming-UniAudio, Ming-UniVision and MingTok from 2025-09-30
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-UniAudio-16B-A3B | speech understanding and generation | 16B total, 3B active (from the repository name) | 2025-09-30 | apache-2.0 | Download | |
| Ming-UniAudio-16B-A3B-Edit | speech editing | 16B total, 3B active (from the repository name) | 2025-09-30 | apache-2.0 | Ming-Lite-Omni (card: editing model 'built upon Ming-Lite-Omni') | Download |
| MingTok-Audio | continuous speech tokenizer | not stated on the card | 2025-09-29 | apache-2.0 | Download | |
| Ming-UniVision-16B-A3B | image understanding and generation | 16B total, 3B active (from the repository name) | 2025-09-30 | apache-2.0 | Download | |
| MingTok-Vision | continuous vision tokenizer | not stated on the card | 2025-09-30 | apache-2.0 | Download |
Ming-lite-omni v1 / v1.5 and Ming-Lite-Uni from 2025-05-02
Card update logs date the Ming-lite-omni preview to 2025-05-04, v1 to 2025-05-28 and v1.5 to 2025-07-15.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ming-Lite-Omni | omni (any-to-any) | 2.8B active (total not stated on the card) | 2025-05-28 | mit | Ling-lite | Download |
| Ming-Lite-Omni-1.5 | omni (any-to-any) | 20.3B total, 3B active in the MoE section | 2025-07-15 | mit | Ling-lite-1.5 | Download · Ming-lite-omni-1.5-FP8 |
| Ming-Lite-Uni | unified image understanding and generation | not stated on the card | 2025-05-02 | mit | Download |
Ring 2 (Ring-2.0, 2.5, 2.6) from 2025-09-08
inclusionAI's reasoning line on the Ling 2 base models. Ring-2.6-1T (May 2026) is the latest Ring release; there is no Ring 3.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ring-mini-2.0 | reasoning | 16B total, 1.4B active | 2025-09-08 | mit | inclusionAI/Ling-mini-base-2.0-20T | Download · Ring-mini-2.0-GGUF |
| Ring-flash-2.0 | reasoning | 100B total, 6.1B active | 2025-09-19 | mit | inclusionAI/Ling-flash-base-2.0 | Download |
| Ring-mini-linear-2.0 | reasoning (hybrid linear attention) | 16.4B total, 1.6B active | 2025-09-24 | mit | inclusionAI/Ling-mini-base-2.0-20T | Download · Ring-mini-linear-2.0-GPTQ-int4 |
| Ring-flash-linear-2.0 | reasoning (hybrid linear attention) | 6.1B active (total not stated in the card text) | 2025-09-25 | mit | inclusionAI/Ling-flash-base-2.0 | Download · Ring-flash-linear-2.0-GPTQ-int4 |
| Ring-flash-linear-2.0-128k | reasoning (hybrid linear attention, 128K context) | 6.1B active (total not stated in the card text) | 2025-10-21 | mit | inclusionAI/Ling-flash-base-2.0 | Download |
| Ring-1T | reasoning | 1T total, 50B active | 2025-10-10 | mit | Ling-1T-base | Download |
| Ring-mini-sparse-2.0-exp | reasoning (sparse attention, experimental) | not stated on the card | 2025-10-14 | mit | inclusionAI/Ling-mini-base-2.0-20T | Download |
| Ring-2.5-1T | reasoning (hybrid linear attention) | trillion-parameter (card) | 2026-02-10 | mit | Download | |
| Ring-2.6-1T | reasoning | trillion-parameter (card) | 2026-05-14 | mit | Download |
Ling 2 (Ling-2.0, 2.5, 2.6) from 2025-09-08
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ling-mini-2.0 | instruct | 16B total, 1.4B active | 2025-09-08 | mit | inclusionAI/Ling-mini-base-2.0 | Download · Ling-mini-2.0-GGUF |
| Ling-mini-base-2.0 | base | 16B total, 1.4B active | 2025-09-08 | mit | Download | |
| Ling-mini-base-2.0-5T | base (5T-token checkpoint) | 16B total, 1.4B active | 2025-09-08 | mit | Download | |
| Ling-mini-base-2.0-10T | base (10T-token checkpoint) | 16B total, 1.4B active | 2025-09-08 | mit | Download | |
| Ling-mini-base-2.0-15T | base (15T-token checkpoint) | 16B total, 1.4B active | 2025-09-08 | mit | Download | |
| Ling-mini-base-2.0-20T | base (20T-token checkpoint) | 16B total, 1.4B active | 2025-09-08 | mit | Download | |
| Ling-flash-2.0 | instruct | 100B total, 6.1B active | 2025-09-17 | mit | inclusionAI/Ling-flash-base-2.0 | Download |
| Ling-flash-base-2.0 | base | 100B total, 6.1B active | 2025-09-17 | mit | Download | |
| Ling-1T | instruct | 1T total, about 50B active | 2025-10-02 | mit | Download | |
| Ling-2.5-1T | instruct | 1T total, 63B active | 2026-02-14 | mit | Download | |
| Ling-2.6-flash | instruct | 104B total, 7.4B active | 2026-04-28 | mit | Download · Ling-2.6-flash-fp8 · Ling-2.6-flash-int4 | |
| Ling-2.6-flash-base | base | about 104B total, about 7.4B active | 2026-06-02 | mit | Download | |
| Ling-2.6-1T | instruct | trillion-parameter (card) | 2026-04-29 | mit | Download | |
| Ling-2.6-1T-base | base | about 1T total, about 63B active | 2026-06-02 | mit | Download |
Ring-lite (Ring 1) from 2025-03-29
The publisher gives these no major version number; they are the reasoning editions built on Ling 1 / Ling 1.5.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ring-lite-distill-preview | reasoning (preview) | 16.8B total, 2.75B active | 2025-03-29 | mit | Ling-lite | Download |
| Ring-lite-linear-preview | reasoning (hybrid linear attention, preview) | 17.1B total, 3.0B active | 2025-04-24 | mit | inclusionAI/Ling-lite | Download |
| Ring-lite | reasoning | 16.8B total, 2.75B active | 2025-06-17 | mit | Ling-lite-1.5 | Download |
| Ring-lite-2506 | reasoning | 16.8B total, 2.75B active | 2025-08-06 | mit | Ling-lite-1.5 | Download |
| Ring-lite-2507 | reasoning | 16.8B total, 2.75B active | 2025-08-01 | mit | inclusionAI/Ling-lite-base-1.5 | Download |
Ling 1 (Ling-lite, Ling-plus, Ling-Coder-lite, Ling-lite-1.5) from 2025-02-28
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ling-lite | instruct | 16.8B total, 2.75B active | 2025-02-28 | mit | Download | |
| Ling-lite-base | base | 16.8B total, 2.75B active | 2025-02-28 | mit | Download | |
| Ling-plus | instruct | 290B total, 28.8B active | 2025-02-28 | mit | Download | |
| Ling-plus-base | base | 290B total, 28.8B active | 2025-02-28 | mit | Download | |
| Ling-Coder-lite | instruct (code) | 16.8B total, 2.75B active | 2025-03-08 | mit | inclusionAI/Ling-Coder-lite-base | Download · Ling-Coder-lite-GPTQ-Int8 |
| Ling-Coder-lite-base | base (code) | 16.8B total, 2.75B active | 2025-03-08 | mit | Download | |
| Ling-lite-1.5 | instruct | 16.8B total, 2.75B active | 2025-05-11 | mit | Download | |
| Ling-lite-base-1.5 | base | 16.8B total, 2.75B active | 2025-05-11 | mit | Download | |
| Ling-lite-1.5-2506 | instruct | 16.8B total, 2.75B active | 2025-07-23 | mit | Download | |
| Ling-lite-1.5-2507 | instruct | 16.8B total, 2.75B active | 2025-08-08 | mit | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Ling-3.0-tiny 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
InternScience (Shanghai AI Laboratory)
Publishes from huggingface.co/InternScience. Checked 2026-10-21.
Agents-A1 from 2026-06-22
InternScience's agent models. Neither repository declares a base in its metadata. Apache-2.0. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Agents-A1 | reasoning (long-horizon agent work) | 35B total, 3B active (MoE; 35,107,181,936 by the Hub's safetensors count) | 2026-06-22 | apache-2.0 | Qwen/Qwen3.5-35B-A3B (not declared; the technical report, arXiv 2606.30616, says it is fine-tuned from Qwen3.5-35B-A3B) | Download · Agents-A1-FP8 · Agents-A1-F16-GGUF · Agents-A1-Q8_0-GGUF · Agents-A1-Q4_K_M-GGUF |
| Agents-A1-4B Agents-A1-4B is InternScience's open-weight dense vision-language model for long-horizon agent work (search, research, instruction following and tool calling), about 4.5B parameters in the Qwen3.5 architecture with a 262,144-token context, published in BF16, FP8 and GGUF under Apache 2.0. | reasoning (long-horizon agent work, vision-language) | 4B (dense; 4,539,265,536 by the Hub's safetensors count) | 2026-07-14 | apache-2.0 | Qwen/Qwen3.5-4B (not declared; inferred from an identical config and matching tensor names, less the 15 MTP tensors) | Our listing (C) · Download · Agents-A1-4B-FP8 · Agents-A1-4B-F16-GGUF · Agents-A1-4B-Q8_0-GGUF · Agents-A1-4B-Q4_K_M-GGUF |
Older versions (5 families, 6 models)
Agents-K1 from 2026-05-06
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Agents-K1 | agent model | not stated (4,411,424,256 by the Hub's safetensors count) | 2026-05-06 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
InternAgent from 2025-05-09
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| InternAgent (No licence in metadata; the Hub reports no parameter count. The card links arXiv 2505.16938.) | agent framework release | not stated | 2025-05-09 | none in the card's metadata | not declared (Qwen2VLForConditionalGeneration in config) | Download |
StructTable from 2024-07-26
Table-image-to-markup models.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| StructTable-base | image-to-text (tables) | not stated (323,502,720 by the Hub's safetensors count) | 2024-07-26 | apache-2.0 | not declared (Pix2Struct architecture) | Download |
| StructTable-InternVL2-1B | image-to-text (tables) | 1B (938,158,976 by the Hub's safetensors count) | 2024-10-18 | apache-2.0 | not declared (InternVL architecture; named after InternVL2-1B) | Download |
ChartVLM from 2024-01-30
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ChartVLM-large (No licence in metadata; no safetensors count.) | chart understanding | not stated | 2024-01-30 | none in the card's metadata | not declared | Download |
MLExpert from 2026-10-01
A separate, later line (machine-learning engineering agent), not an earlier version.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MLExpert | reasoning (ML engineering agent) | 35B total, 3B active (MoE; 34,660,610,688 by the Hub's safetensors count) | 2026-10-01 | apache-2.0 | Qwen/Qwen3.6-35B-A3B (not declared in the metadata; the card's text names it as the base) | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Agents-A1-4B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Jackrong
Publishes from huggingface.co/Jackrong. Checked 2026-10-21.
Qwopus3 from 2026-03-30
Jackrong's Qwopus line: reasoning fine-tunes of Qwen3.x models, trained on distilled traces (the cards cite Claude and GPT among the sources). Qwopus3.5 (from March 2026), Qwopus3.6 (from April) and Qwopus3.8 (from September) follow Qwen's 3.5, 3.6 and 3.8 and are one major family by the decimal-is-minor rule, as Qwen3.5 to Qwen3.8 are one Qwen3 family. Within a generation Jackrong numbers its own revisions (v1, v2, v3, v3.5, Flash, Flash-V2, Coder). Jackrong is a personal Hugging Face account, not an organisation. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwopus3.5-9B-v3 | reasoning (distilled) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-03-30 | apache-2.0 | unsloth/Qwen3.5-9B (Unsloth's copy of Qwen/Qwen3.5-9B) | Download · MLX-Qwopus3.5-9B-v3-4bit · MLX-Qwopus3.5-9B-v3-8bit · Qwopus3.5-9B-v3-GGUF |
| Qwopus3.5-27B-v3 | reasoning (distilled) | 27B (27,356,728,560 by the Hub's safetensors count) | 2026-04-01 | apache-2.0 | unsloth/Qwen3.5-27B (Unsloth's copy of Qwen/Qwen3.5-27B) | Download · MLX-Qwopus3.5-27B-v3-4bit · Qwopus3.5-27B-v3-FP8-vllm-ready · Qwopus3.5-27B-v3-GGUF |
| Qwopus3.5-4B-v3 | reasoning (distilled) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-04-01 | apache-2.0 | unsloth/Qwen3.5-4B (Unsloth's copy of Qwen/Qwen3.5-4B) | Download · Qwopus3.5-4B-v3-GGUF · Qwopus3.5-4B-v3-MTP-GGUF |
| Qwopus3.5-0.8B-v3 | reasoning (distilled) | 0.8B (873,438,784 by the Hub's safetensors count) | 2026-04-12 | apache-2.0 | unsloth/Qwen3.5-0.8B (Unsloth's copy of Qwen/Qwen3.5-0.8B) | Download · Qwopus3.5-0.8B-v3-GGUF |
| Qwopus3.5-27B-v3.5 | reasoning (distilled) | 27B (27,356,728,560 by the Hub's safetensors count) | 2026-04-12 | apache-2.0 | unsloth/Qwen3.5-27B (Unsloth's copy of Qwen/Qwen3.5-27B) | Download · Qwopus3.5-27B-v3.5-GGUF |
| Qwopus3.5-2B-v3 | reasoning (distilled) | 2B (2,274,069,824 by the Hub's safetensors count) | 2026-04-12 | apache-2.0 | unsloth/Qwen3.5-2B (Unsloth's copy of Qwen/Qwen3.5-2B) | Download · Qwopus3.5-2B-v3-GGUF |
| Qwopus3.5-9B-v3.5 | reasoning (distilled) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-04-14 | apache-2.0 | unsloth/Qwen3.5-9B (Unsloth's copy of Qwen/Qwen3.5-9B) | Download |
| Qwopus-GLM-18B-Healed | merge (two 9B models stacked to 18B, then healed) | 18B (16,572,666,096 by the Hub's safetensors count) | 2026-04-20 | apache-2.0 | Jackrong/Qwopus3.5-9B-v3.5 + Jackrong/Qwen3.5-9B-GLM5.1-Distill-v1 | Download · Qwopus-GLM-18B-Merged-GGUF |
| Qwopus3.6-27B-v1-preview | reasoning (preview) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-04-22 | apache-2.0 | Qwen/Qwen3.6-27B | Download · Qwopus3.6-27B-v1-preview-GGUF |
| Qwopus3.6-35B-A3B-v1 | reasoning (distilled) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-05-06 | apache-2.0 | unsloth/Qwen3.6-35B-A3B (Unsloth's copy of Qwen/Qwen3.6-35B-A3B) | Download · Qwopus3.6-35B-A3B-v1-GGUF · Qwopus3.6-35B-A3B-v1-MTP-GGUF |
| Qwopus3.6-27B-v2 | reasoning (distilled) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-05-14 | apache-2.0 | Qwen/Qwen3.6-27B (the metadata writes qwen/ in lower case) | Download · Qwopus3.6-27B-v2-FP8 · Qwopus3.6-27B-v2-GGUF · Qwopus3.6-27B-v2-MLX-4bit · Qwopus3.6-27B-v2-MTP-GGUF |
| Qwopus3.5-9B-Coder | coding (reasoning) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-05-15 | apache-2.0 | Jackrong/Qwopus3.5-9B-v3.5 (from Unsloth's copy of Qwen/Qwen3.5-9B) | Download · Qwopus3.5-9B-Coder-GGUF · Qwopus3.5-9B-Coder-MTP-GGUF |
| Qwopus3.5-4B-Coder | coding (reasoning) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-05-26 | apache-2.0 | Qwen/Qwen3.5-4B | Download · Qwopus3.5-4B-Coder-GGUF · Qwopus3.5-4B-Coder-MTP-GGUF |
| Qwopus3.6-27B-Coder | coding (reasoning) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-06-01 | apache-2.0 | Jackrong/Qwopus3.6-27B-v2 (from Qwen/Qwen3.6-27B) | Download · Qwopus3.6-27B-Coder-Compat-MTP-GGUF · Qwopus3.6-27B-Coder-FP8 · Qwopus3.6-27B-Coder-GGUF · Qwopus3.6-27B-Coder-MTP-GGUF |
| Qwopus3.6-35B-A3B-Coder | coding (reasoning) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-06-21 | apache-2.0 | Jackrong/Qwopus3.6-35B-A3B-v1 + unsloth/Qwen3.6-35B-A3B (Unsloth's copy of Qwen/Qwen3.6-35B-A3B) | Download · Qwopus3.6-35B-A3B-Coder-FP8 · Qwopus3.6-35B-A3B-Coder-MTP-GGUF |
| Qwopus3.6-27B-Fusion-BF16 (Published as a BF16 GGUF; licence other (license_name qwen) in its metadata.) | reasoning (GGUF only) | 27B (27,320,697,856 by the Hub's GGUF count) | 2026-07-30 | other (qwen) | Qwen/Qwen3.6-27B (its metadata calls it quantised from Qwen3.6-27B) | Download · Qwopus3.6-27B-Fusion-GGUF |
| Qwopus3.8-27B-Flash | reasoning (shorter thinking, for agent loops) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-03 | apache-2.0 | Qwen/Qwen3.8-27B | Download · Qwopus3.8-27B-Flash-GGUF |
| Qwopus3.8-27B-Flash-V2 Qwopus3.8-27B-Flash-V2 is Jackrong's open-weight 27B fine-tune of Qwen3.8-27B, by way of his Qwopus3.8-27B-Flash, trained for shorter reasoning in agent loops, published in BF16 and GGUF with a multi-token prediction head under Apache 2.0. (Jackrong's own GGUF repository carries the downloads (769,100 in 30 days against 875 for the BF16 repository).) | reasoning (shorter thinking, for agent loops) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-21 | apache-2.0 | Jackrong/Qwopus3.8-27B-Flash (itself a fine-tune of Qwen/Qwen3.8-27B) | Our listing (C) · Download · Qwopus3.8-27B-Flash-V2-GGUF |
| Qwopus3-5-9B-Coder-Flash (Named 'Qwopus3-5' (with a hyphen) but a Qwen3.5-9B Qwopus, released 1 October 2026.) | coding (reasoning) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-10-01 | apache-2.0 | unsloth/Qwen3.5-9B (Unsloth's copy of Qwen/Qwen3.5-9B) | Download · Qwopus3-5-9B-Coder-Flash-GGUF |
Older versions (13 families, 53 models)
Negentropy from 2026-05-07
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Negentropy-claude-opus-4.7-4B | reasoning (Claude Opus 4.7 traces, per the name) | 4B (4,539,265,536 by the Hub's safetensors count) | 2026-05-07 | apache-2.0 | Qwen/Qwen3.5-4B | Download · Negentropy-claude-opus-4.7-4B-GGUF |
| Negentropy-claude-opus-4.7-9B | reasoning (Claude Opus 4.7 traces, per the name) | 9B (9,409,813,744 by the Hub's safetensors count) | 2026-05-07 | apache-2.0 | Qwen/Qwen3.5-9B | Download · Negentropy-claude-opus-4.7-9B-GGUF |
Trace-Inverter from 2026-05-05
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Trace-Inverter-4B | trace inversion (question answering) | 4B (4,154,588,672 by the Hub's safetensors count) | 2026-05-05 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
DeepSeek-V4 distills from 2026-04-29
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3.5-9B-DeepSeek-V4-Flash | reasoning (distilled from DeepSeek V4, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-04-29 | apache-2.0 | unsloth/Qwen3.5-9B (Unsloth's copy of Qwen/Qwen3.5-9B) | Download · MLX-Qwen3.5-9B-DeepSeek-V4-Flash-4bit · Qwen3.5-9B-DeepSeek-V4-Flash-GGUF · Qwen3.5-9B-DeepSeek-V4-Flash-MTP-GGUF |
| DeepSeek-V4-Pro-Qwen3.5-9B | reasoning (distilled from DeepSeek V4, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-07-19 | apache-2.0 | unsloth/Qwen3.5-9B (Unsloth's copy of Qwen/Qwen3.5-9B) | Download · DeepSeek-V4-Pro-Qwen3.5-9B-MTP-GGUF |
| DeepSeek-V4-Pro-Qwen3.5-4B | reasoning (distilled from DeepSeek V4, per the name) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-07-24 | apache-2.0 | unsloth/Qwen3.5-4B (Unsloth's copy of Qwen/Qwen3.5-4B) | Download · DeepSeek-V4-Pro-Qwen3.5-4B-MTP-GGUF |
GLM5.1-Distill from 2026-04-15
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3.5-9B-GLM5.1-Distill-v1 | reasoning (distilled from GLM-5.1, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-04-15 | apache-2.0 | Qwen/Qwen3.5-9B (not declared; the metadata names the repository itself as its base, and the files have Qwen3.5-9B's architecture and parameter count) | Download · MLX-Qwen3.5-9B-GLM5.1-Distill-v1-6bit · Qwen3.5-9B-GLM5.1-Distill-v1-GGUF |
| Qwen3.5-27B-GLM5.1-Distill-v1 | reasoning (distilled from GLM-5.1, per the name) | 27B (27,356,728,560 by the Hub's safetensors count) | 2026-04-17 | apache-2.0 | unsloth/Qwen3.5-27B (Unsloth's copy of Qwen/Qwen3.5-27B) | Download · Qwen3.5-27B-GLM5.1-Distill-v1-GGUF |
Gemopus-4 from 2026-04-08
Jackrong's Gemma 4 counterpart to Qwopus.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Gemopus-4-E4B-it | reasoning (Opus-style distill on Gemma 4) | E4B (effective size; 7,996,156,490 by the Hub's safetensors count) | 2026-04-08 | apache-2.0 | google/gemma-4-E4B-it (the metadata says google/gemma4-E4B-it, which isn't a repository) | Download · Gemopus-4-E4B-it-GGUF |
| Gemopus-4-26B-A4B-it | reasoning (Opus-style distill on Gemma 4) | 26B total, 4B active (MoE; 26,544,131,376 by the Hub's safetensors count) | 2026-04-09 | apache-2.0 | google/gemma-4-26B-A4B-it (the metadata says google/gemma4-26B-it, which isn't a repository) | Download · Gemopus-4-26B-A4B-it-GGUF |
| Gemopus-4-31B-it | reasoning (Opus-style distill on Gemma 4) | 31B (32,682,372,656 by the Hub's safetensors count) | 2026-04-15 | apache-2.0 | google/gemma-4-31B-it | Download · Gemopus-4-31B-it-GGUF |
Neo from 2026-03-22
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3.5-4B-Neo | reasoning (fine-tune) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-03-22 | apache-2.0 | Qwen/Qwen3.5-4B | Download · Qwen3.5-4B-Neo-GGUF |
| Qwen3.5-9B-Neo | reasoning (fine-tune) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-03-24 | apache-2.0 | Qwen/Qwen3.5-9B | Download · Qwen3.5-9B-Neo-GGUF |
Gemini-3.1-Pro-Reasoning-Distill from 2026-03-10
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3.5-27B-Gemini-3.1-Pro-Reasoning-Distill | reasoning (distilled from Gemini 3.1 Pro traces, per the name) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-03-10 | apache-2.0 | Qwen/Qwen3.5-27B | Download · Qwen3.5-27B-Gemini-3.1-Pro-Reasoning-Distill-GGUF |
| Qwen3.5-9B-Gemini-3.1-Pro-Reasoning-Distill | reasoning (distilled from Gemini 3.1 Pro traces, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-03-10 | apache-2.0 | Qwen/Qwen3.5-9B | Download · Qwen3.5-9B-Gemini-3.1-Pro-Reasoning-Distill-GGUF |
Claude-4.6-Opus-Reasoning-Distilled from 2026-02-27
Named after the Qwen base they fine-tune, with v2 refreshes in March 2026.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Qwen3-4B-2507-Claude-4.6-Opus-Reasoning-Distilled-GGUF (Published only as GGUF, with no licence in its metadata.) | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 4B (4,022,468,096 by the Hub's GGUF count) | 2026-02-27 | none in the card's metadata | Qwen/Qwen3-4B-Thinking-2507 (not declared; inferred from the name and the GGUF file names) | Download |
| Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-02-27 | apache-2.0 | Qwen/Qwen3.5-27B | Download · MLX-Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-4bit · Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 2B (2,274,069,824 by the Hub's safetensors count) | 2026-03-03 | apache-2.0 | Qwen/Qwen3.5-2B (not declared; inferred from the name, architecture and parameter count) | Download · Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-03-03 | apache-2.0 | Qwen/Qwen3.5-4B | Download · MLX-Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-4bit · MLX-Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-6bit · MLX-Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-8bit · Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-03-03 | apache-2.0 | Qwen/Qwen3.5-9B | Download · MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-4bit · MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-6bit · MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-8bit · Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-0.8B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 0.8B (873,438,784 by the Hub's safetensors count) | 2026-03-05 | apache-2.0 | Qwen/Qwen3.5-0.8B | Download · MLX-Qwen3.5-0.8B-Claude-4.6-Opus-Reasoning-Distilled-6bit · Qwen3.5-0.8B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-03-07 | apache-2.0 | Qwen/Qwen3.5-35B-A3B (the metadata writes qwen/ in lower case) | Download · MLX-Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-4bit · MLX-Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-8bit · Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF |
| Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2 | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 4B (4,659,865,088 by the Hub's safetensors count) | 2026-03-16 | apache-2.0 | Qwen/Qwen3.5-4B | Download · Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF |
| Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2 | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-03-16 | apache-2.0 | Qwen/Qwen3.5-9B | Download · MLX-Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-4bit · Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF |
| Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 | reasoning (distilled from Claude Opus 4.6 traces, per the name) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-03-19 | apache-2.0 | Qwen/Qwen3.5-27B | Download · MLX-Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-4bit · MLX-Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-6bit · MLX-Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-8bit · Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF |
DASD from 2026-01-30
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| DASD-1.7B-stage2 | reasoning (distillation experiments) | 1.7B (1,720,574,976 by the Hub's safetensors count) | 2026-01-30 | none in the card's metadata | Qwen/Qwen3-1.7B | Download · DASD-1.7B-stage2-GGUF |
| DASD-4B-Thinking-2507-GRPO-v2 | reasoning (distillation experiments) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-02-10 | apache-2.0 | unsloth/Qwen3-4B-Thinking-2507 (Unsloth's copy of Qwen/Qwen3-4B-Thinking-2507) | Download |
| DASD-4B-Thinking-2507-stage2 | reasoning (distillation experiments) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-02-14 | apache-2.0 | Qwen/Qwen3-4B-Thinking-2507 | Download |
| Qwen3-4B-Thinking-2507-GLM-4.7-Distilled | reasoning (distillation experiments) | 4B (4,022,468,096 by the Hub's safetensors count) | 2026-02-23 | apache-2.0 | Jackrong/DASD-4B-Thinking-2507-GRPO-v2 | Download |
Thinking from 2025-12-10
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama3.2-3B-thinking | reasoning (fine-tune) | 3B (3,212,749,824 by the Hub's safetensors count) | 2025-12-10 | apache-2.0 | unsloth/llama-3.2-3b-instruct-unsloth-bnb-4bit (Unsloth's copy of meta-llama/Llama-3.2-3B-Instruct) | Download |
| Llama3.1-8B-Thinking-R1 | reasoning (fine-tune) | 8B (8,030,261,248 by the Hub's safetensors count) | 2025-12-20 | llama3.1 | unsloth/Llama-3.1-8B-Instruct (Unsloth's copy of meta-llama/Llama-3.1-8B-Instruct) | Download · Llama3.1-8B-Thinking-R1-GGUF |
| Qwen3-0.6B-Thinking | reasoning (fine-tune) | 0.6B (596,049,920 by the Hub's safetensors count) | 2026-01-08 | apache-2.0 | unsloth/qwen3-0.6b-unsloth-bnb-4bit (Unsloth's copy of Qwen/Qwen3-0.6B) | Download · Qwen3-0.6B-Thinking-GGUF |
GPT distills from 2025-09-24
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gpt-oss-120b-Distill-Phi-4-14B | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 14B (14,659,507,200 by the Hub's safetensors count) | 2025-09-24 | apache-2.0 | microsoft/phi-4 | Download · gpt-oss-120b-Distill-Phi-4-14B-GGUF |
| gpt-oss-120b-Distill-Llama3.1-8B-v1 | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 8B (8,030,261,248 by the Hub's safetensors count) | 2025-10-03 | llama3.1 | meta-llama/Llama-3.1-8B | Download |
| gpt-oss-120b-Distill-Llama3.1-8B-v2 | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 8B (8,030,261,248 by the Hub's safetensors count) | 2025-10-03 | llama3.1 | Jackrong/gpt-oss-120b-Distill-Llama3.1-8B-v1 | Download · gpt-oss-120b-Distill-Llama3.1-8B-v2-GGUF |
| gpt-oss-120b-Distill-Qwen3-1.7B-Thinking | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 1.7B (1,720,574,976 by the Hub's safetensors count) | 2025-11-20 | apache-2.0 | Qwen/Qwen3-1.7B | Download · gpt-oss-120b-Distill-Qwen3-1.7B-Thinking-GGUF |
| gpt-oss-120b-Distill-Qwen3-1.7B-Thinking-test1 (A test checkpoint.) | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 1.7B (1,720,574,976 by the Hub's safetensors count) | 2025-11-20 | apache-2.0 | unsloth/qwen3-1.7b-unsloth-bnb-4bit (Unsloth's copy of Qwen/Qwen3-1.7B) | Download · gpt-oss-120b-Distill-Qwen3-1.7B-Thinking-test1-GGUF |
| GPT-5-Distill-Qwen3-4B-Instruct | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 4B (4,022,468,096 by the Hub's safetensors count) | 2025-11-21 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download · GPT-5-Distill-Qwen3-4B-Instruct-GGUF |
| gpt-oss-120b-Distill-Qwen3-4B-Thinking | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 4B (4,022,468,096 by the Hub's safetensors count) | 2025-11-21 | apache-2.0 | Qwen/Qwen3-4B | Download · gpt-oss-120b-Distill-Qwen3-4B-Thinking-GGUF |
| gpt-oss-120b-Distill-Llama3.1-8B-v3 | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 8B (8,030,261,248 by the Hub's safetensors count) | 2025-11-22 | llama3.1 | meta-llama/Llama-3.1-8B (the metadata says Llama3.1-8B without an organisation) | Download |
| GPT-5-Distill-llama3.1-8B-Instruct | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 8B (8,030,261,248 by the Hub's safetensors count) | 2025-11-28 | llama3.1 | meta-llama/Llama-3.1-8B-Instruct | Download |
| GPT-5-Distill-llama3.2-3B-Instruct | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 3B (3,212,749,824 by the Hub's safetensors count) | 2025-11-29 | llama3.2 | unsloth/Llama-3.2-3B-Instruct (Unsloth's copy of meta-llama/Llama-3.2-3B-Instruct) | Download · GPT-5-Distill-llama3.2-3B-Instruct-GGUF |
| GPT-Distill-Qwen3-8B-Thinking | reasoning (distilled from gpt-oss-120b or GPT-5, per the name) | 8B (8,190,735,360 by the Hub's safetensors count) | 2025-11-30 | apache-2.0 | Qwen/Qwen3-8B | Download |
Soren from 2025-09-14
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Soren-Logos-3B (Published only as GGUF.) | instruct (chat, finance) | 3B (3,212,749,888 by the Hub's GGUF count) | 2025-09-14 | llama3 | Jackrong/llama-3.2-3B-Chinese-Elite-v2 + Jackrong/Soren-Oracle-Chat-3B | Download |
| Soren-Oracle-Chat-3B (Published only as GGUF.) | instruct (chat, finance) | 3B (3,212,749,888 by the Hub's GGUF count) | 2025-09-14 | llama3 | meta-llama/Llama-3.2-3B-Instruct (the metadata says llama-3.2-3b-Instruct without an organisation) | Download |
| Soren-Qwen3-4B-Instruct-Finance-v1 | instruct (chat, finance) | 4B (4,022,468,096 by the Hub's safetensors count) | 2025-09-16 | apache-2.0 | Qwen/Qwen3-4B-Instruct-2507 | Download |
Elite from 2025-08-28
Jackrong's first releases (August and September 2025), some only as GGUF.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| llama-3.2-3B-Chinese-Elite (Published only as GGUF.) | instruct (fine-tune) | 3B (3,212,749,888 by the Hub's GGUF count) | 2025-08-28 | apache-2.0 | meta-llama/Llama-3.2-3B-Instruct | Download |
| llama-3.2-3B-Elite (Published only as GGUF.) | instruct (fine-tune) | 3B (3,212,749,888 by the Hub's GGUF count) | 2025-08-28 | apache-2.0 | meta-llama/Llama-3.2-3B-Instruct | Download |
| Llama-3.1-8B-Instruct-Chinese-Elite (Published only as GGUF.) | instruct (fine-tune) | 8B (8,030,261,312 by the Hub's GGUF count) | 2025-08-30 | llama3 | meta-llama/Llama-3.1-8B-Instruct | Download |
| Llama-3.1-8B-Instruct-Elite (Published only as GGUF.) | instruct (fine-tune) | 8B (8,030,261,312 by the Hub's GGUF count) | 2025-08-30 | llama3 | meta-llama/Llama-3.1-8B-Instruct | Download |
| Llama3.3-70B-Instruct-Elite-v1 | instruct (fine-tune) | 70B (70,553,706,496 by the Hub's safetensors count) | 2025-09-06 | llama3 | meta-llama/Llama-3.3-70B-Instruct | Download |
| llama-3.2-3B-Chinese-Elite-v2 | instruct (fine-tune) | 3B (3,212,749,824 by the Hub's safetensors count) | 2025-09-13 | apache-2.0 | unsloth/llama-3.2-3b-instruct-unsloth-bnb-4bit (Unsloth's copy of meta-llama/Llama-3.2-3B-Instruct) | Download |
| llama-3.2-3B-Chinese-Elite-v2-GRPO | instruct (fine-tune) | 3B (3,212,749,824 by the Hub's safetensors count) | 2025-09-14 | apache-2.0 | Jackrong/llama-3.2-3B-Chinese-Elite-v2 | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Qwopus3.8-27B-Flash-V2 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Jina AI (Elastic)
Publishes from huggingface.co/jinaai. Checked 2026-10-21.
jina-embeddings-v5 (text and omni) from 2026-01-22
Jina's fifth embedding generation: v5-text nano (on EuroBERT-210m) and small (on Qwen3-0.6B-Base), January 2026, with per-task repositories (retrieval, text-matching, clustering, classification); and v5-omni nano and small (text, image and audio), created 31 March and 1 April 2026. Per the batch research notes, omni-small was released publicly on 8 May 2026 and reuses the frozen v5-text-small text tower with a Qwen3.5-2B vision encoder and a Qwen2.5-Omni audio encoder, training only the projectors; its card declares no base_model, so none is set. All cc-by-nc-4.0 (non-commercial). Jina's own GGUF and MLX repositories are builds. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-embeddings-v5-omni-small jina-embeddings-v5-omni-small is Jina AI's open-weight embedding model for text, images, video, audio and PDFs, about 1.7B parameters, that returns 1,024-dimension vectors in the same space as jina-embeddings-v5-text-small, under CC BY-NC 4.0 (non-commercial). | embedding (text, image and audio; task adapters included) | 1.63B (Hub safetensors count; the name states none) | 2026-05-08 | cc-by-nc-4.0 | Our listing (D) · Download · jina-embeddings-v5-omni-small-mlx | |
| jina-embeddings-v5-omni-nano | embedding (text, image and audio; task adapters included) | 986M (Hub safetensors count; the name states none) | 2026-04-01 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-nano-mlx | |
| jina-embeddings-v5-omni-small-text-matching | embedding | 1.56B (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-small-text-matching-mlx · jina-embeddings-v5-omni-small-text-matching-GGUF | |
| jina-embeddings-v5-omni-small-retrieval | embedding | 1.56B (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-small-retrieval-mlx · jina-embeddings-v5-omni-small-retrieval-GGUF | |
| jina-embeddings-v5-omni-small-clustering | embedding | 1.56B (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-small-clustering-mlx · jina-embeddings-v5-omni-small-clustering-GGUF | |
| jina-embeddings-v5-omni-small-classification | embedding | 1.56B (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-small-classification-mlx · jina-embeddings-v5-omni-small-classification-GGUF | |
| jina-embeddings-v5-omni-nano-text-matching | embedding | 948M (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-nano-text-matching-mlx · jina-embeddings-v5-omni-nano-text-matching-GGUF | |
| jina-embeddings-v5-omni-nano-retrieval | embedding | 948M (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-nano-retrieval-mlx · jina-embeddings-v5-omni-nano-retrieval-GGUF | |
| jina-embeddings-v5-omni-nano-clustering | embedding | 948M (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-nano-clustering-mlx · jina-embeddings-v5-omni-nano-clustering-GGUF | |
| jina-embeddings-v5-omni-nano-classification | embedding | 948M (Hub safetensors count; the name states none) | 2026-03-31 | cc-by-nc-4.0 | Download · jina-embeddings-v5-omni-nano-classification-mlx · jina-embeddings-v5-omni-nano-classification-GGUF | |
| jina-embeddings-v5-text-small-text-matching | embedding | 596M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-small-text-matching-mlx · jina-embeddings-v5-text-small-text-matching-GGUF | |
| jina-embeddings-v5-text-small-clustering | embedding | 596M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-small-clustering-mlx · jina-embeddings-v5-text-small-clustering-GGUF | |
| jina-embeddings-v5-text-small-classification | embedding | 596M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-small-classification-mlx · jina-embeddings-v5-text-small-classification-GGUF | |
| jina-embeddings-v5-text-nano-text-matching | embedding | 212M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-nano-text-matching-GGUF · jina-embeddings-v5-text-nano-text-matching-mlx | |
| jina-embeddings-v5-text-nano-retrieval | embedding | 212M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-nano-retrieval-GGUF · jina-embeddings-v5-text-nano-retrieval-mlx | |
| jina-embeddings-v5-text-nano-clustering | embedding | 212M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-nano-clustering-GGUF · jina-embeddings-v5-text-nano-clustering-mlx | |
| jina-embeddings-v5-text-nano-classification | embedding | 212M (Hub safetensors count; the name states none) | 2026-02-10 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-nano-classification-GGUF · jina-embeddings-v5-text-nano-classification-mlx | |
| jina-embeddings-v5-text-small-retrieval | embedding | 596M (Hub safetensors count; the name states none) | 2026-02-05 | cc-by-nc-4.0 | Download · jina-embeddings-v5-text-small-retrieval-mlx · jina-embeddings-v5-text-small-retrieval-GGUF | |
| jina-embeddings-v5-text-small | embedding | 596M (Hub safetensors count; the name states none) | 2026-01-22 | cc-by-nc-4.0 | Qwen/Qwen3-0.6B-Base | Download · jina-embeddings-v5-text-small-mlx |
| jina-embeddings-v5-text-nano | embedding | 212M (Hub safetensors count; the name states none) | 2026-01-22 | cc-by-nc-4.0 | EuroBERT/EuroBERT-210m | Download · jina-embeddings-v5-text-nano-mlx |
Older versions (10 families, 32 models)
jina-vlm from 2025-11-17
Jina's earlier embedding, reranker and reader models, newest first. Several are fine-tunes of Qwen models, declared in their cards (base).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-vlm | instruct (vision-language) | 2.48B (Hub safetensors count; the name states none) | 2025-11-17 | cc-by-nc-4.0 | Qwen/Qwen3-1.7B-Base | Download · jina-vlm-mlx |
jina-code-embeddings from 2025-08-19
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-code-embeddings-1.5b | embedding | 1.5B (Hub safetensors count 1.54B) | 2025-08-26 | cc-by-nc-4.0 | Qwen/Qwen2.5-Coder-1.5B | Download · jina-code-embeddings-1.5b-GGUF · jina-code-embeddings-1.5b-mlx |
| jina-code-embeddings-0.5b | embedding | 0.5B (Hub safetensors count 494M) | 2025-08-19 | cc-by-nc-4.0 | Qwen/Qwen2.5-Coder-0.5B | Download · jina-code-embeddings-0.5b-GGUF · jina-code-embeddings-0.5b-mlx |
jina-embeddings-v4 from 2025-05-07
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-embeddings-v4-vllm-text-matching | embedding | 3.75B (Hub safetensors count; the name states none) | 2025-07-01 | none in the card metadata | Download | |
| jina-embeddings-v4-vllm-retrieval | embedding | 3.75B (Hub safetensors count; the name states none) | 2025-07-01 | none in the card metadata | Download | |
| jina-embeddings-v4-vllm-code | embedding | 3.75B (Hub safetensors count; the name states none) | 2025-07-01 | none in the card metadata | Download | |
| jina-embeddings-v4 | embedding | 3.75B (Hub safetensors count; the name states none) | 2025-05-07 | none in the card metadata | Download · jina-embeddings-v4-text-matching-GGUF · jina-embeddings-v4-text-code-GGUF · jina-embeddings-v4-text-retrieval-GGUF · jina-embeddings-v4-mlx-8bit |
ReaderLM from 2024-09-06
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ReaderLM-v2 | instruct (HTML to Markdown and JSON) | 1.54B (Hub safetensors count; the name states none) | 2025-01-13 | cc-by-nc-4.0 | Download | |
| reader-lm-1.5b | instruct (HTML to Markdown) | 1.5B (Hub safetensors count 1.54B) | 2024-09-06 | cc-by-nc-4.0 | Download | |
| reader-lm-0.5b | instruct (HTML to Markdown) | 0.5B (Hub safetensors count 494M) | 2024-09-06 | cc-by-nc-4.0 | Download |
jina-embeddings-v3 from 2024-09-05
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-embeddings-v3 | embedding | 572M (Hub safetensors count; the name states none) | 2024-09-05 | cc-by-nc-4.0 | Download · jina-embeddings-v3-hf |
jina-clip from 2024-05-21
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-clip-v2 | embedding | 865M (Hub safetensors count; the name states none) | 2024-10-08 | cc-by-nc-4.0 | Download | |
| jina-clip-v1 | embedding (text and image, CLIP) | 223M (Hub safetensors count; the name states none) | 2024-05-21 | apache-2.0 | Download |
jina-reranker from 2024-04-15
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-reranker-v3.5 | reranker | 597M (Hub safetensors count; the name states none) | 2026-07-14 | cc-by-nc-4.0 | Qwen/Qwen3-0.6B | Download · jina-reranker-v3.5-GGUF · jina-reranker-v3.5-mlx |
| jina-reranker-v3 | reranker | 597M (Hub safetensors count; the name states none) | 2025-09-18 | cc-by-nc-4.0 | Qwen/Qwen3-0.6B | Download · jina-reranker-v3-mlx · jina-reranker-v3-GGUF |
| jina-reranker-m0 | reranker (multimodal) | 2.44B (Hub safetensors count; the name states none) | 2025-03-27 | cc-by-nc-4.0 | Qwen/Qwen2-VL-2B-Instruct | Download · jina-reranker-m0-GGUF |
| jina-reranker-v2-base-multilingual | reranker (multilingual) | 278M (Hub safetensors count; the name states none) | 2024-06-19 | cc-by-nc-4.0 | Download | |
| jina-reranker-v1-turbo-en | reranker | 38M (Hub safetensors count; the name states none) | 2024-04-15 | apache-2.0 | Download | |
| jina-reranker-v1-tiny-en | reranker | 33M (Hub safetensors count; the name states none) | 2024-04-15 | apache-2.0 | Download |
jina-colbert from 2024-01-23
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-colbert-v2-64 | embedding (late interaction, ColBERT, 64 dimensions) | 559M (Hub safetensors count; the name states none) | 2024-08-22 | cc-by-nc-4.0 | Download | |
| jina-colbert-v2 | embedding (late interaction, ColBERT) | 559M (Hub safetensors count; the name states none) | 2024-08-16 | cc-by-nc-4.0 | Download | |
| jina-colbert-v1-en | embedding (late interaction, ColBERT) | 137M (Hub safetensors count; the name states none) | 2024-01-23 | apache-2.0 | Download |
jina-embeddings-v2 from 2023-09-27
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-embeddings-v2-base-es | embedding | 161M (Hub safetensors count; the name states none) | 2024-01-24 | apache-2.0 | Download | |
| jina-embeddings-v2-base-de | embedding | 161M (Hub safetensors count; the name states none) | 2024-01-12 | apache-2.0 | Download | |
| jina-embeddings-v2-base-zh | embedding | 161M (Hub safetensors count; the name states none) | 2024-01-10 | apache-2.0 | Download | |
| jina-embeddings-v2-base-code | embedding | 161M (Hub safetensors count; the name states none) | 2023-11-17 | apache-2.0 | Download | |
| jina-embeddings-v2-small-en | embedding | 33M (Hub safetensors count; the name states none) | 2023-09-27 | apache-2.0 | Download | |
| jina-embeddings-v2-base-en | embedding | 137M (Hub safetensors count; the name states none) | 2023-09-27 | apache-2.0 | Download |
jina-embedding v1 from 2023-07-06
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| jina-embedding-t-en-v1 | embedding | not stated (no safetensors count on the Hub) | 2023-07-21 | apache-2.0 | Download | |
| jina-embedding-l-en-v1 | embedding | not stated (no safetensors count on the Hub) | 2023-07-09 | apache-2.0 | Download | |
| jina-embedding-b-en-v1 | embedding | not stated (no safetensors count on the Hub) | 2023-07-07 | apache-2.0 | Download | |
| jina-embedding-s-en-v1 | embedding | not stated (no safetensors count on the Hub) | 2023-07-06 | apache-2.0 | Download |
k2-fsa (Next-gen Kaldi)
Publishes from huggingface.co/k2-fsa, github.com/k2-fsa. Checked 2026-10-21.
OmniVoice from 2026-03-30
Zero-shot multilingual text-to-speech built on Qwen/Qwen3-0.6B as its language-model backbone (declared in the card). Licence: the card's metadata named apache-2.0 until 3 July 2026 and names none now; its text says the pre-trained model is CC-BY-NC (no version) because of its training data, the code is Apache 2.0, and the audio tokenizer folder carries the Boson Higgs Audio 2 Community License (all per the batch research notes). OmniVoice-Emilia's card was not read. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| OmniVoice OmniVoice is k2-fsa's open-weight zero-shot text-to-speech model for over 600 languages, initialised from Qwen3-0.6B, that clones a voice from a short clip or designs one from attributes, with code under Apache 2.0 and weights under CC-BY-NC by the card. | text-to-speech (zero-shot voice cloning) | 613M (Hub safetensors count; the name states none) | 2026-03-30 | CC-BY-NC for the weights per the card text (no version given; no licence in the metadata since 3 July 2026, apache-2.0 before); audio tokenizer under the Boson Higgs Audio 2 Community License; code Apache 2.0 | Qwen/Qwen3-0.6B | Our listing (C) · Download |
| OmniVoice-Emilia | text-to-speech (per its name, trained on Emilia; card not read) | 613M (Hub safetensors count; the name states none) | 2026-03-30 | none in the card metadata | Download |
Older versions (2 families, 10 models)
ZipVoice from 2025-06-15
Flow-matching text-to-speech (ZipVoice and ZipVoice-Dialog in one repository, PyTorch, safetensors and ONNX files). No licence in the metadata.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ZipVoice | text-to-speech (ZipVoice and ZipVoice-Dialog folders) | not stated (no safetensors count on the Hub) | 2025-06-15 | none in the card metadata | Download |
Zipformer ASR (icefall and sherpa-onnx exports) from 2023-12-12
Speech recognisers trained with icefall and exported for sherpa-onnx. Most k2-fsa models are published from maintainers' personal accounts (for example csukuangfj), not this organisation; only what k2-fsa itself hosts is listed. Audio-tagging models, sherpa bundles and wheels, TTS evaluation models and the Flow2GAN vocoder are left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| icefall-asr-zipformer-wenetspeech-streaming-small | speech-to-text | not stated (no safetensors count on the Hub) | 2024-08-21 | apache-2.0 | Download | |
| icefall-asr-zipformer-wenetspeech-streaming-large | speech-to-text | not stated (no safetensors count on the Hub) | 2024-08-21 | apache-2.0 | Download | |
| icefall-asr-zipformer-wenetspeech-small | speech-to-text | not stated (no safetensors count on the Hub) | 2024-08-21 | apache-2.0 | Download | |
| icefall-asr-zipformer-wenetspeech-large | speech-to-text | not stated (no safetensors count on the Hub) | 2024-08-21 | apache-2.0 | Download | |
| sherpa-onnx-zipformer-korean-2024-06-24 | speech-to-text | not stated (no safetensors count on the Hub) | 2024-06-24 | none in the card metadata | Download | |
| sherpa-onnx-streaming-zipformer-korean-2024-06-16 | speech-to-text | not stated (no safetensors count on the Hub) | 2024-06-16 | none in the card metadata | Download | |
| sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13 | speech-to-text | not stated (no safetensors count on the Hub) | 2023-12-13 | apache-2.0 | Download · sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-int8-2023-12-13 · sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-fp16-2023-12-13 | |
| sherpa-onnx-zipformer-gigaspeech-2023-12-12 | speech-to-text | not stated (no safetensors count on the Hub) | 2023-12-12 | apache-2.0 | Download | |
| sherpa-onnx-streaming-zipformer-multi-zh-hans-2023-12-12 | speech-to-text | not stated (no safetensors count on the Hub) | 2023-12-12 | apache-2.0 | Download |
Liquid AI
Publishes from huggingface.co/LiquidAI. Checked 2026-10-21.
LFM2 (LFM2 and LFM2.5) from 2025-07-10
Liquid's hybrid (convolution and attention) models: LFM2 from July 2025 and LFM2.5 from December 2025, treated as one major version as Qwen3.5 sits in Qwen3. Sizes 230M to 24B-A2B, with vision (VL), audio, embedding, ColBERT, encoder and task fine-tunes (Extract, RAG, Tool, Math, translation, Japanese, PII, Longevity). Licence lfm1.0 (Liquid's LFM Open Licence, which conditions commercial use on revenue under $10 million, per our liquid-d1 listing). Liquid's GGUF, ONNX, MLX and FP8 repositories are builds; its DSpark speculative-decoding drafters and two smoltalk LoRA adapters are left out. LFM2.5-2.6B's date is Liquid's blog post of 4 August 2026, per the batch research notes. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| LFM2.5-2.6B LFM2.5-2.6B is Liquid AI's open-weight 2.6B hybrid model of convolution and attention layers for on-device agents, with tool calling, thinking always on and a 131,072-token context, published on 4 August 2026 with Liquid's own GGUF, MLX and ONNX builds. | instruct | 2.6B (Hub safetensors count 2.70B) | 2026-08-04 | other (lfm1.0) | Our listing (D) · Download · LFM2.5-2.6B-GGUF · LFM2.5-2.6B-ONNX · LFM2.5-2.6B-MLX · LFM2.5-2.6B-MLX-bf16 · LFM2.5-2.6B-MLX-8bit · LFM2.5-2.6B-MLX-6bit · LFM2.5-2.6B-MLX-5bit · LFM2.5-2.6B-MLX-4bit · LFM2.5-2.6B-MLX-mxfp8 · LFM2.5-2.6B-MLX-mxfp4 · LFM2.5-2.6B-MLX-nvfp4 | |
| LFM2.5-24B-A2B | instruct | 24B total, 2B active (MoE) (Hub safetensors count 23.97B) | 2026-10-12 | other (lfm1.0) | Download · LFM2.5-24B-A2B-ONNX · LFM2.5-24B-A2B-GGUF · LFM2.5-24B-A2B-MLX-4bit · LFM2.5-24B-A2B-MLX-5bit · LFM2.5-24B-A2B-MLX-6bit · LFM2.5-24B-A2B-MLX-8bit · LFM2.5-24B-A2B-MLX-bf16 · LFM2.5-24B-A2B-FP8 | |
| LFM2.5-24B-A2B-Base | base | 24B total, 2B active (MoE) (Hub safetensors count 23.97B) | 2026-10-09 | other (lfm1.0) | Download | |
| LFM2.5-VL-3B | instruct (vision-language) | 3B (Hub safetensors count 3.12B) | 2026-08-11 | other (lfm1.0) | Download · LFM2.5-VL-3B-ONNX · LFM2.5-VL-3B-MLX-4bit · LFM2.5-VL-3B-MLX-6bit · LFM2.5-VL-3B-MLX-5bit · LFM2.5-VL-3B-MLX-8bit · LFM2.5-VL-3B-MLX-bf16 · LFM2.5-VL-3B-GGUF | |
| LFM2.5-2.6B-Base | base | 2.6B (Hub safetensors count 2.70B) | 2026-08-01 | other (lfm1.0) | Download | |
| LFM2.5-Encoder-350M | encoder (masked language model) | 350M (Hub safetensors count 354M) | 2026-07-27 | other (lfm1.0) | Download · LFM2.5-Encoder-350M-GGUF | |
| LFM2.5-Encoder-230M | encoder (masked language model) | 230M (Hub safetensors count 230M) | 2026-07-27 | other (lfm1.0) | Download · LFM2.5-Encoder-230M-GGUF | |
| LFM2.5-Encoder-350M-Prompt-Router | classifier | 350M (Hub safetensors count 355M) | 2026-07-13 | other (lfm1.0) | Download | |
| LFM2.5-Encoder-350M-Policy-Linter | classifier | 350M (Hub safetensors count 355M) | 2026-07-11 | other (lfm1.0) | Download | |
| LFM2.5-Encoder-350M-Diffusion | diffusion language model (per name; card not read) | 350M (Hub safetensors count 354M) | 2026-07-10 | other (lfm1.0) | Download | |
| LFM2.5-Encoder-350M-Spellchecker | classifier | 350M (Hub safetensors count 357M) | 2026-07-04 | other (lfm1.0) | Download | |
| LFM2.5-230M | instruct | 230M (Hub safetensors count 230M) | 2026-06-24 | other (lfm1.0) | Download · LFM2.5-230M-GGUF · LFM2.5-230M-MLX-bf16 · LFM2.5-230M-MLX-8bit · LFM2.5-230M-MLX-6bit · LFM2.5-230M-MLX-5bit · LFM2.5-230M-MLX-4bit · LFM2.5-230M-ONNX | |
| LFM2.5-230M-Base | base | 230M (Hub safetensors count 230M) | 2026-06-16 | other (lfm1.0) | Download | |
| LFM2.5-Encoder-350M-PII-Detector | classifier | 350M (Hub safetensors count 355M) | 2026-06-12 | other (lfm1.0) | Download | |
| LFM2.5-8B-A1B-Base | base | 8B total, 1B active (MoE) (Hub safetensors count 8.47B) | 2026-05-28 | other (lfm1.0) | Download | |
| LFM2.5-8B-A1B | instruct | 8B total, 1B active (MoE) (Hub safetensors count 8.47B) | 2026-05-28 | other (lfm1.0) | Download · LFM2.5-8B-A1B-GGUF · LFM2.5-8B-A1B-ONNX · LFM2.5-8B-A1B-MLX-4bit · LFM2.5-8B-A1B-MLX-6bit · LFM2.5-8B-A1B-MLX-5bit · LFM2.5-8B-A1B-MLX-bf16 · LFM2.5-8B-A1B-MLX-8bit | |
| LFM2.5-VL-450M-Extract | instruct (vision-language) | 450M (Hub safetensors count 449M) | 2026-05-26 | other (lfm1.0) | Download · LFM2.5-VL-450M-Extract-GGUF | |
| LFM2.5-VL-1.6B-Extract | instruct (vision-language) | 1.6B (Hub safetensors count 1.60B) | 2026-05-26 | other (lfm1.0) | Download · LFM2.5-VL-1.6B-Extract-GGUF | |
| LFM2.5-Audio-1.5B-JP | speech and audio (audio in, text or audio out) | 1.5B (Hub safetensors count 1.47B) | 2026-05-26 | other (lfm1.0) | Download · LFM2.5-Audio-1.5B-JP-GGUF | |
| LFM2.5-1.2B-JP-202606 | instruct | 1.2B (Hub safetensors count 1.17B) | 2026-05-26 | other (lfm1.0) | Download · LFM2.5-1.2B-JP-202606-GGUF · LFM2.5-1.2B-JP-202606-MLX-8bit · LFM2.5-1.2B-JP-202606-MLX-6bit · LFM2.5-1.2B-JP-202606-MLX-4bit · LFM2.5-1.2B-JP-202606-MLX-5bit · LFM2.5-1.2B-JP-202606-ONNX | |
| LFM2-2.6B-Longevity | instruct | 2.6B (Hub safetensors count 2.57B) | 2026-05-25 | other (lfm1.0) | Download | |
| LFM2-1.2B-Longevity | instruct | 1.2B (Hub safetensors count 1.17B) | 2026-05-25 | other (lfm1.0) | Download | |
| LFM2.5-ColBERT-350M | embedding (late interaction, ColBERT) | 350M (Hub safetensors count 353M) | 2026-05-20 | other (lfm1.0) | Download · LFM2.5-ColBERT-350M-GGUF | |
| LFM2.5-Embedding-350M | embedding | 350M (Hub safetensors count 354M) | 2026-05-05 | other (lfm1.0) | Download · LFM2.5-Embedding-350M-GGUF | |
| LFM2.5-VL-450M | instruct (vision-language) | 450M (Hub safetensors count 449M) | 2026-04-08 | other (lfm1.0) | Download · LFM2.5-VL-450M-GGUF · LFM2.5-VL-450M-ONNX · LFM2.5-VL-450M-MLX-8bit · LFM2.5-VL-450M-MLX-bf16 · LFM2.5-VL-450M-MLX-6bit · LFM2.5-VL-450M-MLX-5bit · LFM2.5-VL-450M-MLX-4bit | |
| LFM2.5-350M-Base | base | 350M (Hub safetensors count 354M) | 2026-03-31 | other (lfm1.0) | Download | |
| LFM2.5-350M | instruct | 350M (Hub safetensors count 354M) | 2026-03-31 | other (lfm1.0) | Download · LFM2.5-350M-GGUF · LFM2.5-350M-ONNX · LFM2.5-350M-MLX-bf16 · LFM2.5-350M-MLX-8bit · LFM2.5-350M-MLX-6bit · LFM2.5-350M-MLX-5bit · LFM2.5-350M-MLX-4bit | |
| LFM2-24B-A2B | instruct | 24B total, 2B active (MoE) (Hub safetensors count 23.84B) | 2026-02-24 | other (lfm1.0) | Download · LFM2-24B-A2B-GGUF · LFM2-24B-A2B-ONNX · LFM2-24B-A2B-MLX-bf16 · LFM2-24B-A2B-MLX-8bit · LFM2-24B-A2B-MLX-6bit · LFM2-24B-A2B-MLX-5bit · LFM2-24B-A2B-MLX-4bit | |
| LFM2.5-1.2B-Thinking | reasoning | 1.2B (Hub safetensors count 1.17B) | 2026-01-20 | other (lfm1.0) | Download · LFM2.5-1.2B-Thinking-ONNX · LFM2.5-1.2B-Thinking-GGUF · LFM2.5-1.2B-Thinking-MLX-4bit · LFM2.5-1.2B-Thinking-MLX-5bit · LFM2.5-1.2B-Thinking-MLX-6bit · LFM2.5-1.2B-Thinking-MLX-8bit · LFM2.5-1.2B-Thinking-MLX-bf16 | |
| LFM2.5-1.2B-Instruct | instruct | 1.2B (Hub safetensors count 1.17B) | 2026-01-06 | other (lfm1.0) | Download · LFM2.5-1.2B-Instruct-GGUF · LFM2.5-1.2B-Instruct-ONNX · LFM2.5-1.2B-Instruct-MLX-4bit · LFM2.5-1.2B-Instruct-MLX-5bit · LFM2.5-1.2B-Instruct-MLX-6bit · LFM2.5-1.2B-Instruct-MLX-8bit · LFM2.5-1.2B-Instruct-MLX-bf16 | |
| LFM2.5-VL-1.6B | instruct (vision-language) | 1.6B (Hub safetensors count 1.60B) | 2026-01-05 | other (lfm1.0) | Download · LFM2.5-VL-1.6B-GGUF · LFM2.5-VL-1.6B-ONNX | |
| LFM2.5-1.2B-Base | base | 1.2B (Hub safetensors count 1.17B) | 2026-01-05 | other | Download · LFM2.5-1.2B-Base-GGUF · LFM2.5-1.2B-Base-ONNX | |
| LFM2-2.6B-Transcript | instruct | 2.6B (Hub safetensors count 2.57B) | 2026-01-05 | other (lfm1.0) | Download · LFM2-2.6B-Transcript-GGUF · LFM2-2.6B-Transcript-ONNX | |
| LFM2.5-1.2B-JP | instruct | 1.2B (Hub safetensors count 1.17B) | 2026-01-04 | other (lfm1.0) | Download · LFM2.5-1.2B-JP-GGUF · LFM2.5-1.2B-JP-ONNX · LFM2.5-1.2B-JP-MLX-4bit · LFM2.5-1.2B-JP-MLX-5bit · LFM2.5-1.2B-JP-MLX-6bit · LFM2.5-1.2B-JP-MLX-8bit · LFM2.5-1.2B-JP-MLX-bf16 | |
| LFM2-2.6B-Exp | instruct (experimental, RL) | 2.6B (Hub safetensors count 2.57B) | 2025-12-25 | other (lfm1.0) | Download · LFM2-2.6B-Exp-GGUF | |
| LFM2.5-Audio-1.5B | speech and audio (audio in, text or audio out) | 1.5B (Hub safetensors count 1.47B) | 2025-12-18 | other (lfm1.0) | Download · LFM2.5-Audio-1.5B-GGUF-LEAP · LFM2.5-Audio-1.5B-GGUF · LFM2.5-Audio-1.5B-ONNX | |
| LFM2-ColBERT-350M | embedding (late interaction, ColBERT) | 350M (Hub safetensors count 353M) | 2025-10-28 | other (lfm1.0) | Download · LFM2-ColBERT-350M-GGUF | |
| LFM2-VL-3B | instruct (vision-language) | 3B (Hub safetensors count 3.00B) | 2025-10-22 | other (lfm1.0) | Download · LFM2-VL-3B-GGUF | |
| LFM2-8B-A1B | instruct | 8B total, 1B active (MoE) (Hub safetensors count 8.34B) | 2025-10-07 | other (lfm1.0) | Download · LFM2-8B-A1B-GGUF · LFM2-8B-A1B-ONNX | |
| LFM2-350M-PII-Extract-JP | instruct | 350M (Hub safetensors count 354M) | 2025-09-30 | other (lfm1.0) | Download · LFM2-350M-PII-Extract-JP-GGUF | |
| LFM2-2.6B | instruct | 2.6B (Hub safetensors count 2.57B) | 2025-09-22 | other (lfm1.0) | Download · LFM2-2.6B-GGUF | |
| LFM2-350M-Extract | instruct | 350M (Hub safetensors count 354M) | 2025-09-03 | other (lfm1.0) | Download · LFM2-350M-Extract-GGUF | |
| LFM2-350M-ENJP-MT | instruct | 350M (Hub safetensors count 354M) | 2025-09-03 | other (lfm1.0) | Download · LFM2-350M-ENJP-MT-GGUF | |
| LFM2-1.2B-Tool | instruct | 1.2B (Hub safetensors count 1.17B) | 2025-09-03 | other (lfm1.0) | Download · LFM2-1.2B-Tool-GGUF | |
| LFM2-1.2B-RAG | instruct | 1.2B (Hub safetensors count 1.17B) | 2025-09-03 | other (lfm1.0) | Download · LFM2-1.2B-RAG-GGUF | |
| LFM2-Audio-1.5B | speech and audio (audio in, text or audio out) | 1.5B (Hub safetensors count 1.47B) | 2025-08-28 | other (lfm1.0) | Download · LFM2-Audio-1.5B-GGUF | |
| LFM2-350M-Math | instruct | 350M (Hub safetensors count 354M) | 2025-08-25 | other (lfm1.0) | Download · LFM2-350M-Math-GGUF | |
| LFM2-1.2B-Extract | instruct | 1.2B (Hub safetensors count 1.17B) | 2025-08-22 | other (lfm1.0) | Download · LFM2-1.2B-Extract-GGUF | |
| LFM2-VL-450M | instruct (vision-language) | 450M (Hub safetensors count 451M) | 2025-08-12 | other (lfm1.0) | Download · LFM2-VL-450M-GGUF | |
| LFM2-VL-1.6B | instruct (vision-language) | 1.6B (Hub safetensors count 1.58B) | 2025-08-12 | other (lfm1.0) | Download · LFM2-VL-1.6B-GGUF | |
| LFM2-700M | instruct | 700M (Hub safetensors count 742M) | 2025-07-10 | other (lfm1.0) | Download · LFM2-700M-GGUF | |
| LFM2-350M | instruct | 350M (Hub safetensors count 354M) | 2025-07-10 | other (lfm1.0) | Download · LFM2-350M-GGUF | |
| LFM2-1.2B | instruct | 1.2B (Hub safetensors count 1.17B) | 2025-07-10 | other (lfm1.0) | Download · LFM2-1.2B-GGUF |
d1 from 2026-10-05
Liquid's open-weight decision models, d1-3B (from LFM2.5-VL-3B) and d1-omni-600M (from LFM2.5-Encoder-350M), 5 October 2026; graded together in our liquid-d1 listing, which also covers the hosted d1 API. Not a numbered generation, so "major" is false.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| d1-omni-600M d1 is Liquid AI's decision model family. It answers typed yes or no, choice and score questions about text and images with probabilities, through the hosted d1 and d1-ultra models and the open-weight d1-3B and d1-omni-600M. | decision model (text with images or audio; experimental per our listing) | 600M (Hub safetensors count 587M) | 2026-10-05 | other (lfm1.0) | LiquidAI/LFM2.5-Encoder-350M | Our listing (E) · Download · d1-omni-600M-GGUF · d1-omni-600M-MLX-fp16 · d1-omni-600M-MLX-8bit |
| d1-3B d1 is Liquid AI's decision model family. It answers typed yes or no, choice and score questions about text and images with probabilities, through the hosted d1 and d1-ultra models and the open-weight d1-3B and d1-omni-600M. | decision model (typed yes or no, choice and score with probabilities; text and images) | 3B (Hub safetensors count 3.12B) | 2026-10-05 | other (lfm1.0) | LiquidAI/LFM2.5-VL-3B | Our listing (E) · Download · d1-3B-GGUF · d1-3B-w8a8 · d1-3B-MLX-8bit · d1-3B-MLX-4bit · d1-3B-MLX-fp16 |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: LFM2.5-2.6B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Meta
Publishes from huggingface.co/meta-models, huggingface.co/meta-llama. Checked 2026-10-21.
Muse from 2026-08-10
Meta's current open-weight line, published under meta-models (verified org 'Meta Inc.'). Muse Glimmer (Apache 2.0, ungated) and Muse Spark Open (Muse Model Licence, gated) are both 'Muse'; treated as one major family as FAMILIES.md names it. Glimmer's date is the launch post (10 Aug 2026), Spark Open's the release (15 Oct 2026), both from our listings; the Hub createdAt dates are 2026-08-09 and 2026-09-15.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Muse Glimmer 30B Muse Glimmer 30B is Meta's open-weight dense multimodal model for local agent work, 29.6B parameters with text and image input and a 128K-token context, published on Hugging Face under Apache 2.0 with GGUF and ExecuTorch builds. | instruct (reasoning, text and image input) | 30B (dense, multimodal) | 2026-08-10 | apache-2.0 | Muse Spark (card: 'distilled from Muse Spark') | Our listing (BB) · Download · Muse-Glimmer-30B-GGUF · Muse-Glimmer-30B-ExecuTorch-PTE |
| Muse Glimmer 30B assistant (DFlash drafter) (Card: the lightweight 'drafter' model for Muse Glimmer 30B, based on DFlash.) | draft model for speculative decoding with Muse Glimmer 30B | 2.56B (Hub safetensors count; the card states no figure) | 2026-08-10 | apache-2.0 | Download | |
| Muse Spark Open 816B-A42B Muse Spark Open is Meta's open-weight multimodal model for coding and agentic work, with 816B total and 42B active parameters and a 1M-token context, released on 15 October 2026 under the Muse Model Licence for self-hosting. (Gated with manual review (form asks for legal name, date of birth, country, affiliation, job title, per our listing). The card itself answered 401 without access, so no base is recorded.) | instruct (reasoning, multimodal) | 816B total, 42B active (MoE) | 2026-10-15 | muse-model-license | Our listing (D) · Download (gated) · Muse-Spark-Open-816B-A42B-NVFP4 |
Older versions (6 families, 50 models)
Llama 4 from 2025-04-01
All meta-llama repositories are gated (manual review, Llama licence acceptance). Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama-4-Scout-17B-16E | base | 109B total, 17B active, 16 experts | 2025-04-02 | llama4 | Download (gated) · Llama-4-Scout-17B-16E-Original | |
| Llama-4-Scout-17B-16E-Instruct | instruct | 109B total, 17B active, 16 experts | 2025-04-02 | llama4 | meta-llama/Llama-4-Scout-17B-16E | Download (gated) · Llama-4-Scout-17B-16E-Instruct-Original |
| Llama-4-Maverick-17B-128E | base | 400B total, 17B active, 128 experts | 2025-04-02 | llama4 | Download (gated) · Llama-4-Maverick-17B-128E-Original | |
| Llama-4-Maverick-17B-128E-Instruct | instruct | 400B total, 17B active, 128 experts | 2025-04-01 | llama4 | meta-llama/Llama-4-Maverick-17B-128E | Download (gated) · Llama-4-Maverick-17B-128E-Instruct-Original · Llama-4-Maverick-17B-128E-Instruct-FP8 · Llama-4-Maverick-17B-128E-Instruct-FP8-Original |
Llama 3 from 2024-04-17
Llama 3, 3.1, 3.2 and 3.3 grouped as one major version by FAMILIES.md's rule. Licences differ by minor version (llama3, llama3.1, llama3.2, llama3.3 community licences). Gated. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Meta-Llama-3-8B | base | 8B | 2024-04-17 | llama3 | Download (gated) | |
| Meta-Llama-3-8B-Instruct | instruct | 8B | 2024-04-17 | llama3 | Download (gated) | |
| Meta-Llama-3-70B | base | 70B | 2024-04-17 | llama3 | Download (gated) | |
| Meta-Llama-3-70B-Instruct | instruct | 70B | 2024-04-17 | llama3 | meta-llama/Meta-Llama-3-70B | Download (gated) |
| Llama-3.1-8B | base | 8B | 2024-07-14 | llama3.1 | Download (gated) | |
| Llama-3.1-8B-Instruct | instruct | 8B | 2024-07-18 | llama3.1 | meta-llama/Meta-Llama-3.1-8B | Download (gated) |
| Llama-3.1-70B | base | 70B | 2024-07-14 | llama3.1 | Download (gated) | |
| Llama-3.1-70B-Instruct | instruct | 70B | 2024-07-16 | llama3.1 | meta-llama/Meta-Llama-3.1-70B | Download (gated) |
| Llama-3.1-405B | base | 405B | 2024-07-16 | llama3.1 | Download (gated) · Llama-3.1-405B-FP8 | |
| Llama-3.1-405B-Instruct | instruct | 405B | 2024-07-16 | llama3.1 | meta-llama/Meta-Llama-3.1-405B | Download (gated) · Llama-3.1-405B-Instruct-FP8 |
| Llama-3.2-1B | base | 1B | 2024-09-18 | llama3.2 | Download (gated) | |
| Llama-3.2-1B-Instruct | instruct | 1B | 2024-09-18 | llama3.2 | Download (gated) · Llama-3.2-1B-Instruct-QLORA_INT4_EO8 · Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8 | |
| Llama-3.2-3B | base | 3B | 2024-09-18 | llama3.2 | Download (gated) | |
| Llama-3.2-3B-Instruct | instruct | 3B | 2024-09-18 | llama3.2 | Download (gated) · Llama-3.2-3B-Instruct-QLORA_INT4_EO8 · Llama-3.2-3B-Instruct-SpinQuant_INT4_EO8 | |
| Llama-3.2-11B-Vision | base (vision) | 11B | 2024-09-18 | llama3.2 | Download (gated) | |
| Llama-3.2-11B-Vision-Instruct | instruct (vision) | 11B | 2024-09-18 | llama3.2 | Download (gated) | |
| Llama-3.2-90B-Vision | base (vision) | 90B | 2024-09-19 | llama3.2 | Download (gated) | |
| Llama-3.2-90B-Vision-Instruct | instruct (vision) | 90B | 2024-09-19 | llama3.2 | Download (gated) | |
| Llama-3.3-70B-Instruct | instruct | 70B | 2024-11-26 | llama3.3 | meta-llama/Llama-3.1-70B | Download (gated) |
Llama 2 from 2023-07-11
Gated. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama-2-7b-hf | base | 7B | 2023-07-13 | llama2 | Download (gated) · Llama-2-7b | |
| Llama-2-7b-chat-hf | chat | 7B | 2023-07-13 | llama2 | Download (gated) · Llama-2-7b-chat | |
| Llama-2-13b-hf | base | 13B | 2023-07-13 | llama2 | Download (gated) · Llama-2-13b | |
| Llama-2-13b-chat-hf | chat | 13B | 2023-07-13 | llama2 | Download (gated) · Llama-2-13b-chat | |
| Llama-2-70b-hf | base | 70B | 2023-07-11 | llama2 | Download (gated) · Llama-2-70b | |
| Llama-2-70b-chat-hf | chat | 70B | 2023-07-14 | llama2 | Download (gated) · Llama-2-70b-chat |
Code Llama from 2024-03-13
Publisher names it Code Llama; built on Llama 2 and under the Llama 2 licence. The meta-llama copies were created on 2024-03-13/14 (Code Llama itself was released in 2023 under another organisation not checked here). Gated. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| CodeLlama-7b-hf | base | 7B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-7b-Python-hf | python | 7B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-7b-Instruct-hf | instruct | 7B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-13b-hf | base | 13B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-13b-Python-hf | python | 13B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-13b-Instruct-hf | instruct | 13B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-34b-hf | base | 34B | 2024-03-14 | llama2 | Download (gated) | |
| CodeLlama-34b-Python-hf | python | 34B | 2024-03-14 | llama2 | Download (gated) | |
| CodeLlama-34b-Instruct-hf | instruct | 34B | 2024-03-14 | llama2 | Download (gated) | |
| CodeLlama-70b-hf | base | 70B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-70b-Python-hf | python | 70B | 2024-03-13 | llama2 | Download (gated) | |
| CodeLlama-70b-Instruct-hf | instruct | 70B | 2024-03-13 | llama2 | Download (gated) |
Llama Guard from 2023-12-05
Meta's safety classifiers, versioned 1 to 4 alongside Llama 2 to Llama 4. Gated. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama Guard 7B | safety classifier | 7B | 2023-12-05 | llama2 | Download (gated) | |
| Llama Guard 2 8B | safety classifier | 8B | 2024-04-17 | llama3 | Download (gated) | |
| Llama Guard 3 8B | safety classifier | 8B | 2024-07-22 | llama3.1 | meta-llama/Meta-Llama-3.1-8B | Download (gated) · Llama-Guard-3-8B-INT8 |
| Llama Guard 3 1B | safety classifier | 1B | 2024-09-20 | llama3.2 | Download (gated) · Llama-Guard-3-1B-INT4 | |
| Llama Guard 3 11B Vision | safety classifier (vision) | 11B | 2024-09-20 | llama3.2 | Download (gated) | |
| Llama Guard 4 12B | safety classifier (multimodal) | 12B | 2025-04-23 | llama4 | Download (gated) |
Prompt Guard from 2024-07-21
Gated. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Prompt Guard 86M | prompt-injection classifier | 86M | 2024-07-21 | llama3.1 | Download (gated) | |
| Llama Prompt Guard 2 22M | prompt-injection classifier | 22M | 2025-04-28 | llama4 | Download (gated) | |
| Llama Prompt Guard 2 86M | prompt-injection classifier | 86M | 2025-04-28 | llama4 | Download (gated) |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Muse Glimmer 30B 5, Muse Spark Open 1. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
MiniMax
Publishes from huggingface.co/MiniMaxAI. Checked 2026-10-21.
MiniMax-M2 from 2025-10-22
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. M2.5's date is our listing's firstReleased (its LICENSE-MODEL says version release date 2026-02-13). All four members state or ship the same 230B/10B shape (Hub safetensors count 228.7B for each); only M2's card states it. M2.5's card metadata says modified-mit but the repository's licence since 6 March 2026 is LICENSE-MODEL, the MiniMax Model License; M2.7 ships a non-commercial licence.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniMax-M2 | reasoning | 230B total, 10B active | 2025-10-22 | modified-mit | Download | |
| MiniMax-M2.1 | reasoning | 230B total, 10B active (Hub count; card doesn't restate) | 2025-12-20 | modified-mit | Download | |
| MiniMax-M2.5 MiniMax-M2.5 is MiniMax's open-weight mixture-of-experts model for coding and agent work, about 229B parameters in FP8 with a 196,608-token context and thinking always on, published on Hugging Face on 12 February 2026. | reasoning | 230B total, 10B active (Hub count; card doesn't restate) | 2026-02-12 | other (MiniMax Model License in LICENSE-MODEL; card metadata still says modified-mit) | Our listing (C) · Download | |
| MiniMax-M2.7 MiniMax-M2.7 is MiniMax's open-weight mixture-of-experts model for coding, agent harnesses and office work, about 229B parameters in FP8 with thinking always on, published on Hugging Face on 9 April 2026 under a non-commercial licence. | reasoning | 230B total, 10B active (Hub count; card doesn't restate) | 2026-04-09 | other (MiniMax non-commercial licence; commercial use needs written authorisation) | Our listing (C) · Download |
MiniMax-M3 from 2026-06-02
Newer than our listed family; no listing yet. MiniMax M3.1 Flash is announced (MiniMax's API lists MiniMax-M3.1-Flash and routes plan requests for M2.5 and M2.7 to it from 19 October 2026) but its weights aren't published: MiniMaxAI's newest repository on 21 October 2026 is MiniMax-Music3. It is queued for a listing in research/open-weights-2026-10-20/BATCH.md once the weights are out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniMax-M3 | instruct | ~428B total, ~23B activated | 2026-06-02 | other (MiniMax Community License, non-commercial) | Download · MiniMax-M3-MXFP8 |
Older versions (3 families, 7 models)
MiniMax-M1 from 2025-06-05
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. The -hf repositories are the publisher's own Transformers-format conversions, listed as builds.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniMax-M1-40k | reasoning | 456B total, 45.9B activated | 2025-06-05 | apache-2.0 | MiniMaxAI/MiniMax-Text-01 | Download · MiniMax-M1-40k-hf |
| MiniMax-M1-80k | reasoning | 456B total, 45.9B activated | 2025-06-13 | apache-2.0 | MiniMaxAI/MiniMax-Text-01 | Download · MiniMax-M1-80k-hf |
MiniMax-01 from 2025-01-15
Dated by LICENSE-MODEL's 'Model Release Date: 15 January 2025' (repositories created 2025-01-12). MiniMax-Text-01-hf is the publisher's Transformers-format conversion.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniMax-Text-01 | instruct | 456B total, 45.9B activated | 2025-01-15 | other (MiniMax Model License; code MIT) | Download · MiniMax-Text-01-hf | |
| MiniMax-VL-01 | instruct | 456B (Hub count; card states no total) | 2025-01-15 | other (MiniMax Model License; code MIT) | Download |
SynLogic from 2025-05-30
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. Research models: Qwen2.5 bases trained with reinforcement learning on MiniMax's SynLogic data; not a MiniMax pretrained family.
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: MiniMax-M2.5 5, MiniMax-M2.7 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Mistral AI
Publishes from huggingface.co/mistralai. Checked 2026-10-21.
Voxtral from 2025-07-01
Mistral's speech line: Voxtral Mini 3B and Small 24B (speech understanding, July 2025), Voxtral 4B TTS (cc-by-nc-4.0), Voxtral Mini 4B Realtime 2602 (streaming transcription, January 2026, built on Ministral-3-3B-Base-2512) and its Arabic fine-tune (October 2026, no licence in the metadata). Mistral names product lines rather than numbered majors, so Voxtral is one family; "major" is false because it isn't a numbered generation. Mistral's API serves the realtime model too (our mistral-voxtral-transcribe listing, per the batch notes). Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Voxtral-Mini-4B-Realtime-2602 Voxtral Mini 4B Realtime 2602 is Mistral AI's open-weight streaming speech recognition model, a causal audio encoder of about 970M parameters on a Ministral 3B decoder, 4.4B in all, transcribing 13 languages live under Apache 2.0. | speech-to-text (streaming, realtime) | 4B (Hub safetensors count 4.43B) | 2026-01-21 | apache-2.0 | mistralai/Ministral-3-3B-Base-2512 | Our listing (C) · Download |
| Voxtral-Mini-4B-Realtime-Arabic | speech-to-text (streaming, Arabic) | 4B (Hub safetensors count 4.43B) | 2026-10-08 | none in the card metadata | mistralai/Ministral-3-3B-Base-2512 | Download |
| Voxtral-4B-TTS-2603 | text-to-speech | 4B | 2025-11-17 | cc-by-nc-4.0 | mistralai/Ministral-3-3B-Base-2512 | Download |
| Voxtral-Small-24B-2507 | speech understanding (audio and text in, text out) | 24B (Hub safetensors count 24.26B) | 2025-07-01 | apache-2.0 | mistralai/Mistral-Small-24B-Base-2501 | Download |
| Voxtral-Mini-3B-2507 | speech understanding (audio and text in, text out) | 3B (Hub safetensors count 4.68B) | 2025-07-01 | apache-2.0 | Download |
Older versions (18 families, 49 models)
Shieldstral from 2026-07-16
Mistral's other open-weight lines. They aren't all older than Voxtral; Mistral names product lines (Ministral, Mistral Small, Magistral, Devstral and so on) rather than one numbered major version, so each line is a family here, newest first.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Shieldstral-1.0-3B | safety classifier (per name; card not read) | 3B (Hub safetensors count 3.85B) | 2026-07-16 | apache-2.0 | mistralai/Ministral-3-3B-Base-2512 | Download |
Mistral Medium 3.5 from 2026-03-31
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Medium-3.5-128B | instruct (card not read) | 128B (Hub safetensors count 127.70B) | 2026-03-31 | other | Download |
Leanstral from 2026-03-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Leanstral-1.5-119B-A6B | instruct (Lean proof assistant, per name; card not read) | 119B total, 6B active (MoE) | 2026-07-01 | apache-2.0 | Download | |
| Leanstral-2603 | instruct (Lean proof assistant, per name; card not read) | not stated (no safetensors count on the Hub) | 2026-03-11 | apache-2.0 | Download |
Mistral Small 4 from 2026-01-23
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Small-4-119B-2603 | instruct (per Mistral, hybrid reasoning; card not read) | 119B (Hub safetensors count 119.40B) | 2026-01-23 | apache-2.0 | Download · Mistral-Small-4-119B-2603-NVFP4 |
Mistral Large 3 from 2025-11-28
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Large-3-675B-Base-2512 | base | 675B (no safetensors count on the Hub) | 2025-11-30 | apache-2.0 | Download | |
| Mistral-Large-3-675B-Instruct-2512 | instruct | 675B (no safetensors count on the Hub) | 2025-11-28 | apache-2.0 | Download · Mistral-Large-3-675B-Instruct-2512-BF16 · Mistral-Large-3-675B-Instruct-2512-NVFP4 |
Devstral 2 from 2025-11-28
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Devstral-Small-2-24B-Instruct-2512 | instruct (agentic coding) | 24B (Hub safetensors count 24.01B) | 2025-11-28 | apache-2.0 | Download | |
| Devstral-2-123B-Instruct-2512 | instruct (agentic coding) | 123B (Hub safetensors count 125.03B) | 2025-11-28 | other | Download |
Ministral 3 from 2025-10-31
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ministral-3-8B-Reasoning-2512 | reasoning | 8B (Hub safetensors count 8.92B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-8B-Reasoning-2512-GGUF | |
| Ministral-3-8B-Instruct-2512 | instruct (vision-language; the main repository holds FP8, BF16 is a separate repository) | 8B (Hub safetensors count 8.92B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-8B-Instruct-2512-BF16 · Ministral-3-8B-Instruct-2512-GGUF | |
| Ministral-3-8B-Base-2512 | base | 8B (Hub safetensors count 8.92B) | 2025-10-31 | apache-2.0 | Download | |
| Ministral-3-3B-Reasoning-2512 | reasoning | 3B (Hub safetensors count 4.25B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-3B-Reasoning-2512-GGUF | |
| Ministral-3-3B-Instruct-2512 | instruct (vision-language; the main repository holds FP8, BF16 is a separate repository) | 3B (Hub safetensors count 3.85B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-3B-Instruct-2512-BF16 · Ministral-3-3B-Instruct-2512-GGUF · Ministral-3-3B-Instruct-2512-ONNX | |
| Ministral-3-3B-Base-2512 | base | 3B (Hub safetensors count 3.85B) | 2025-10-31 | apache-2.0 | Download | |
| Ministral-3-14B-Reasoning-2512 | reasoning | 14B (Hub safetensors count 13.95B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-14B-Reasoning-2512-GGUF | |
| Ministral-3-14B-Instruct-2512 | instruct (vision-language; the main repository holds FP8, BF16 is a separate repository) | 14B (Hub safetensors count 13.95B) | 2025-10-31 | apache-2.0 | Download · Ministral-3-14B-Instruct-2512-BF16 · Ministral-3-14B-Instruct-2512-GGUF | |
| Ministral-3-14B-Base-2512 | base | 14B (Hub safetensors count 13.95B) | 2025-10-31 | apache-2.0 | Download |
Magistral from 2025-06-04
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Magistral-Small-2509 | reasoning | 24.01B (Hub safetensors count; the name states none) | 2025-09-12 | apache-2.0 | mistralai/Mistral-Small-3.2-24B-Instruct-2506 | Download · Magistral-Small-2509-GGUF |
| Magistral-Small-2507 | reasoning | 23.57B (Hub safetensors count; the name states none) | 2025-07-18 | apache-2.0 | mistralai/Mistral-Small-3.1-24B-Instruct-2503 | Download · Magistral-Small-2507-GGUF |
| Magistral-Small-2506 | reasoning | 23.57B (Hub safetensors count; the name states none) | 2025-06-04 | apache-2.0 | mistralai/Mistral-Small-3.1-24B-Instruct-2503 | Download · Magistral-Small-2506_gguf |
Devstral from 2025-05-12
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Devstral-Small-2507 | instruct (agentic coding) | 23.57B (Hub safetensors count; the name states none) | 2025-07-04 | apache-2.0 | mistralai/Mistral-Small-3.1-24B-Instruct-2503 | Download · Devstral-Small-2507_gguf |
| Devstral-Small-2505 | instruct (agentic coding) | 23.57B (Hub safetensors count; the name states none) | 2025-05-12 | apache-2.0 | mistralai/Mistral-Small-3.1-24B-Instruct-2503 | Download · Devstral-Small-2505_gguf |
Mistral Small 3 (3, 3.1, 3.2) from 2025-01-23
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Small-3.2-24B-Instruct-2506 | instruct | 24B (Hub safetensors count 24.01B) | 2025-06-19 | apache-2.0 | Download | |
| Mistral-Small-3.1-24B-Base-2503 | base | 24B (Hub safetensors count 24.01B) | 2025-03-16 | apache-2.0 | Download | |
| Mistral-Small-3.1-24B-Instruct-2503 | instruct | 24B (Hub safetensors count 24.01B) | 2025-03-11 | apache-2.0 | Download | |
| Mistral-Small-24B-Instruct-2501 | instruct | 24B (Hub safetensors count 23.57B) | 2025-01-28 | apache-2.0 | Download | |
| Mistral-Small-24B-Base-2501 | base | 24B (Hub safetensors count 23.57B) | 2025-01-23 | apache-2.0 | Download |
Ministral (2410) from 2024-10-15
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ministral-8B-Instruct-2410 | instruct | 8B (Hub safetensors count 8.02B) | 2024-10-15 | other (mrl) | Download |
Mistral Small (2409) from 2024-09-17
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Small-Instruct-2409 | instruct | 22.25B (Hub safetensors count; the name states none) | 2024-09-17 | other (mrl) | Download |
Pixtral 12B from 2024-09-11
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Pixtral-12B-Base-2409 | base | 12B | 2024-10-17 | apache-2.0 | Download | |
| Pixtral-12B-2409 | instruct | 12B | 2024-09-11 | apache-2.0 | Download |
Mistral Large 2 (2407, 2411) and Pixtral Large from 2024-07-24
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Pixtral-Large-Instruct-2411 | instruct | not stated (no safetensors count on the Hub) | 2024-11-14 | other (mrl) | Download | |
| Mistral-Large-Instruct-2411 | instruct | 122.61B (Hub safetensors count; the name states none) | 2024-11-14 | other (mrl) | Download | |
| Mistral-Large-Instruct-2407 | instruct | 122.61B (Hub safetensors count; the name states none) | 2024-07-24 | other (mrl) | Download (gated) |
Mistral NeMo from 2024-07-17
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-Nemo-Base-2407 | base | 12.25B (Hub safetensors count; the name states none) | 2024-07-18 | apache-2.0 | Download | |
| Mistral-Nemo-Instruct-2407 | instruct | 12.25B (Hub safetensors count; the name states none) | 2024-07-17 | apache-2.0 | Download · Mistral-Nemo-Instruct-FP8-2407 |
Codestral, Mamba-Codestral and Mathstral from 2024-05-29
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mathstral-7B-v0.1 | instruct (maths) | 7B (Hub safetensors count 7.25B) | 2024-07-16 | apache-2.0 | Download | |
| Mamba-Codestral-7B-v0.1 | instruct (code, Mamba architecture) | 7B (Hub safetensors count 7.29B) | 2024-07-16 | apache-2.0 | Download | |
| Codestral-22B-v0.1 | instruct (code) | 22B (Hub safetensors count 22.25B) | 2024-05-29 | other (mnpl) | Download |
Mixtral from 2023-12-01
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mixtral-8x22B-v0.1 | base | 8x22B (Hub safetensors count 140.62B) | 2024-04-16 | apache-2.0 | Download | |
| Mixtral-8x22B-Instruct-v0.1 | instruct | 8x22B (Hub safetensors count 140.63B) | 2024-04-16 | apache-2.0 | Download | |
| Mixtral-8x7B-Instruct-v0.1 | instruct | 8x7B (Hub safetensors count 46.70B) | 2023-12-10 | apache-2.0 | Download | |
| Mixtral-8x7B-v0.1 | base | 8x7B (Hub safetensors count 46.70B) | 2023-12-01 | apache-2.0 | Download |
Mistral 7B from 2023-09-20
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mistral-7B-v0.3 | base | 7B (Hub safetensors count 7.25B) | 2024-05-22 | apache-2.0 | Download | |
| Mistral-7B-Instruct-v0.3 | instruct | 7B (Hub safetensors count 7.25B) | 2024-05-22 | apache-2.0 | Download | |
| Mistral-7B-Instruct-v0.2 | instruct | 7B (Hub safetensors count 7.24B) | 2023-12-11 | apache-2.0 | Download | |
| Mistral-7B-Instruct-v0.1 | instruct | 7B (Hub safetensors count 7.24B) | 2023-09-27 | apache-2.0 | Download | |
| Mistral-7B-v0.1 | base | 7B (Hub safetensors count 7.24B) | 2023-09-20 | apache-2.0 | Download |
Moonshot AI (Kimi)
Publishes from huggingface.co/moonshotai. Checked 2026-10-21.
Kimi K3 from 2026-07-27
Released date is our listing's firstReleased; the Hub repository was created 2026-06-13. Moonshot publishes one repository (no separate base, no first-party quantised build in moonshotai). Card: 2.8T total, 104B activated, native multimodal.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-K3 Kimi K3 is Moonshot AI's open-weight multimodal mixture-of-experts model, 2.8T parameters with 104B active and a 1M-token context, trained in MXFP4 and published on Hugging Face under the modified-MIT Kimi K3 Licence. | reasoning | 2.8T total, 104B activated | 2026-07-27 | other (Kimi K3 License) | Our listing (B) · Download |
Older versions (6 families, 18 models)
Kimi K2 from 2025-07-11
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. K2-Base and K2-Instruct dated 2025-07-11 (Instruct repository creation; Base repository created 2025-07-03). K2.5, K2.6 and K2.7-Code carry the K2 name and share K2's 1T/32B shape, so they sit here (K2.5's card says it was continually pretrained atop Kimi-K2-Base).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-K2-Base | base | 1T total, 32B activated | 2025-07-11 | modified-mit (Moonshot Modified MIT) | Download | |
| Kimi-K2-Instruct | instruct | 1T total, 32B activated | 2025-07-11 | modified-mit (Moonshot Modified MIT) | Download | |
| Kimi-K2-Instruct-0905 | instruct | 1T total, 32B activated | 2025-09-03 | modified-mit (Moonshot Modified MIT) | Download | |
| Kimi-K2-Thinking | reasoning | 1T total, 32B activated | 2025-11-04 | modified-mit (Moonshot Modified MIT) | Download | |
| Kimi-K2.5 | instruct | 1T total, 32B activated | 2026-01-01 | modified-mit (Moonshot Modified MIT) | moonshotai/Kimi-K2-Base | Download |
| Kimi-K2.6 | instruct | 1T total, 32B activated | 2026-04-14 | modified-mit (Moonshot Modified MIT) | Download | |
| Kimi-K2.7-Code | instruct | 1T total, 32B activated | 2026-06-11 | modified-mit (Moonshot Modified MIT) | moonshotai/Kimi-K2.6 | Download |
Kimi Linear from 2025-10-30
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-Linear-48B-A3B-Base | base | 48B total, 3B activated | 2025-10-30 | mit | Download | |
| Kimi-Linear-48B-A3B-Instruct | instruct | 48B total, 3B activated | 2025-10-30 | mit | Download |
Kimi-Dev from 2025-06-16
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. Coding model fine-tuned from Qwen2.5-72B per its card metadata.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-Dev-72B | instruct | 72B | 2025-06-16 | mit | Qwen/Qwen2.5-72B | Download |
Kimi-VL from 2025-04-09
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. Vision-language MoE on Moonlight; MoonViT is its vision encoder, published separately.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-VL-A3B-Instruct | instruct | 16B total, 3B activated (2.8B in the language decoder) | 2025-04-09 | mit | moonshotai/Moonlight-16B-A3B | Download |
| Kimi-VL-A3B-Thinking | reasoning | 16B total, 3B activated | 2025-04-09 | mit | moonshotai/Kimi-VL-A3B-Instruct | Download |
| Kimi-VL-A3B-Thinking-2506 | reasoning | 16B total, 3B activated | 2025-06-21 | mit | moonshotai/Kimi-VL-A3B-Instruct | Download |
| MoonViT-SO-400M | vision encoder | 400M | 2025-04-10 | mit | SigLIP-SO-400M | Download |
Kimi-Audio from 2025-04-25
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release. Audio model; the card says it is based on and modified from Qwen2.5-7B.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Kimi-Audio-7B | base | 7B | 2025-04-25 | mit (Qwen2.5-derived code under Apache-2.0) | Qwen2.5-7B | Download |
| Kimi-Audio-7B-Instruct | instruct | 7B | 2025-04-25 | mit (Qwen2.5-derived code under Apache-2.0) | Qwen2.5-7B | Download |
Moonlight from 2025-02-22
Release dates are the date the publisher states (card news line, licence 'Release Date', or our listing's firstReleased) where noted; otherwise the Hub's repository creation date (createdAt), which can precede the public release.
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Kimi K3 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
NVIDIA (Nemotron)
Publishes from huggingface.co/nvidia. Checked 2026-10-21.
Nemotron 3 from 2025-12-15
Hybrid Mamba-2/Transformer mixture-of-experts line. Nemotron 3.5 (Lightning, Super VL, Bolt) is a minor version inside it, and Nemotron Labs 3 research checkpoints built on Nemotron 3 weights are included. Language and vision-language generators only: Nemotron 3 / 3.5 Content Safety, Embed, Rerank, Parse, ASR and Diarization are left out. Dates are the card's stated Hugging Face release date or our listing's firstReleased where given, otherwise the repository's createdAt from the Hub API (repositories are often created private some days before release: Nano 30B-A3B was created 2025-12-04 and released 2025-12-15). Licence is the Hub licence identifier from the card front matter. Nothing in this family is gated.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Nemotron 3 Nano 30B-A3B Nemotron 3 Nano 30B A3B is NVIDIA's open-weight hybrid Mamba-Transformer mixture of experts for reasoning and agent work, 30B parameters with 3.5B active, published on Hugging Face in BF16, FP8 and NVFP4. | instruct (reasoning on/off) | 30B total, 3B active (31.6B per the safetensors index) | 2025-12-15 | nvidia-nemotron-open-model-license | Our listing (B) · Download · NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 · NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 | |
| Nemotron 3 Nano 30B-A3B Base | base | 30B total, 3B active | 2025-12-03 | nvidia-nemotron-open-model-license | Download | |
| Nemotron 3 Super 120B-A12B Nemotron 3 Super is NVIDIA's open-weight hybrid Mamba-2 and mixture-of-experts reasoning model for agents, 120B parameters with 12B active, published on Hugging Face in BF16, FP8 and NVFP4 under the NVIDIA Nemotron Open Model License. | instruct (reasoning on/off) | 120B total, 12B active | 2026-03-11 | nvidia-nemotron-open-model-license | Our listing (B) · Download · NVIDIA-Nemotron-3-Super-120B-A12B-FP8 · NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 · Nemotron-3-Super-120B-A12B-BF16-MTPv2 | |
| Nemotron 3 Super 120B-A12B Base | base | 120B total, 12B active | 2026-03-04 | nvidia-nemotron-open-model-license | Download | |
| Nemotron 3 Nano 4B Nemotron 3 Nano 4B is NVIDIA's open-weight small hybrid Mamba-2 and Transformer reasoning model for agents on edge devices, 4B parameters compressed from Nemotron Nano 9B v2, published in BF16, FP8 and GGUF under the NVIDIA Nemotron Open Model License. (Compressed from Nemotron Nano 9B v2 with the Nemotron Elastic framework, per the card.) | instruct (reasoning on/off) | 4B | 2026-03-16 | nvidia-nemotron-open-model-license | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | Our listing (C) · Download · NVIDIA-Nemotron-3-Nano-4B-FP8 · NVIDIA-Nemotron-3-Nano-4B-GGUF |
| Nemotron 3 Nano Omni 30B-A3B Reasoning Nemotron 3 Nano Omni is NVIDIA's open-weight multimodal model that takes video, audio, images and text and returns text, a hybrid Mamba-Transformer mixture of experts with about 31B parameters and 3B active, published in BF16, FP8 and NVFP4. | reasoning (multimodal: video, audio, image, text in; text out) | 30B total, 3B active | 2026-04-28 | nvidia-open-model-agreement | Our listing (C) · Download · Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 · Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 | |
| Nemotron 3 Ultra 550B-A55B Nemotron 3 Ultra is NVIDIA's largest open-weight hybrid Mamba-2 and latent mixture-of-experts reasoning model for agents, 550B parameters with 55B active and up to 1M tokens of context, published on Hugging Face in BF16 and NVFP4 under OpenMDW-1.1. | instruct (reasoning on/off) | 550B total, 55B active | 2026-06-04 | openmdw-1.1 | Our listing (B) · Download · NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 | |
| Nemotron 3 Ultra 550B-A55B Base | base | 550B total, 55B active | 2026-06-04 | openmdw-1.1 | Download | |
| Nemotron 3 Ultra 550B-A55B GenRM | generative reward model | 550B total, 55B active | 2026-06-04 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | Download |
| Nemotron 3.5 Lightning 30B-A3B Nemotron 3.5 Lightning is NVIDIA's open-weight hybrid Mamba-2 and mixture-of-experts reasoning model for agents, 30B parameters with 3B active and up to 1M tokens of context, published on Hugging Face in BF16 and NVFP4 under OpenMDW-1.1. | instruct (reasoning on/off) | 30B total, 3B active | 2026-08-11 | openmdw-1.1 | Our listing (B) · Download · NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash · NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark | |
| Nemotron 3.5 Lightning 30B-A3B Base | base | 30B total, 3B active | 2026-08-05 | openmdw-1.1 | Download | |
| Nemotron 3.5 Super VL 120B-A12B Nemotron 3.5 Super VL is NVIDIA's open-weight vision-language model for agents, a 120B mixture of experts with 12B active parameters that reads text, images and video. NVIDIA hosts a free trial endpoint, and the weights are under OpenMDW-1.1. (The card says the language backbone is initialised from Nemotron 3 Super Base. Our listing is data/tools (kind model, with NVIDIA's hosted trial endpoint); these are its weights.) | instruct, vision-language (text, image, video in) | 120B total, 12B active | 2026-10-15 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16 | Our listing (C) · Download · NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 · NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr |
| Nemotron 3.5 Super VL 120B-A12B Base | base, vision-language | 120B total, 12B active | 2026-10-15 | openmdw-1.1 | Download | |
| Nemotron 3.5 Bolt 74B-A9B (Compressed from Nemotron 3.5 Super VL, per the card. The 24GB repositories are NVIDIA's native 2-bit build of it.) | instruct, vision-language | 74B total, 9B active | 2026-10-15 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | Download · NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB · NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF · NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr · NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF |
| Nemotron Labs 3 Elastic 30B-A3B | research (Nemotron Labs, elastic/nested model) | 30B total, 3B active | 2026-05-07 | nvidia-open-model-license | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | Download · NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8 · NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-NVFP4 |
| Nemotron Labs 3 Puzzle 75B-A9B | research (Nemotron Labs, Puzzle-compressed) | 75B total, 9B active | 2026-06-24 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | Download · NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-FP8 · NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 |
| Nemotron 3 Labs Ultra Math SFT | research (Nemotron Labs, maths SFT) | 550B total, 55B active | 2026-09-02 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | Download |
| Nemotron 3 Labs Ultra Math RL | research (Nemotron Labs, maths RL) | 550B total, 55B active | 2026-09-02 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | Download |
| Nemotron Labs 3 Competitive Coding 550B-A55B (Published only in NVFP4; the card says it was fine-tuned on reasoning traces distilled from GLM-5.2.) | research (Nemotron Labs, competitive programming) | 550B total, 55B active | 2026-09-03 | openmdw-1.1 | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | Download |
| nemotron-3.5-asr-streaming-0.6b Nemotron 3.5 ASR Streaming 0.6B is NVIDIA's open-weight streaming speech recognition model, a 600M-parameter FastConformer with an RNN-T decoder, transcribing 32 language-locales with punctuation and an optional detected-language tag at chunk sizes from 80 ms to 1.12 s, under OpenMDW-1.1. | speech recognition (cache-aware streaming, multilingual) | 0.6B (638M per the safetensors count) | 2026-05-15 | openmdw-1.1 | Our listing (C) · Download |
Parakeet from 2023-12-27
NVIDIA's FastConformer speech-recognition line, published with NeMo. NVIDIA gives it no major version: single models are versioned (parakeet-tdt-0.6b v2, v3), so every parakeet-* repository in NVIDIA's organisation is here. Released is the Hub createdAt; parameters are the size in the name. None is gated. Licences are cc-by-4.0 up to parakeet-tdt-0.6b-v3 and the NVIDIA Open Model License for the Danish, end-of-utterance, multitalker, Vietnamese and unified models.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| parakeet-rnnt-1.1b | speech recognition (RNN-T, English) | 1.1B | 2023-12-27 | cc-by-4.0 | Download | |
| parakeet-ctc-1.1b | speech recognition (CTC, English) | 1.1B | 2023-12-28 | cc-by-4.0 | Download | |
| parakeet-rnnt-0.6b | speech recognition (RNN-T, English) | 0.6B | 2023-12-28 | cc-by-4.0 | Download | |
| parakeet-ctc-0.6b | speech recognition (CTC, English) | 0.6B | 2023-12-28 | cc-by-4.0 | Download | |
| parakeet-tdt-1.1b | speech recognition (TDT, English) | 1.1B | 2024-01-25 | cc-by-4.0 | Download | |
| parakeet-tdt_ctc-1.1b | speech recognition (hybrid TDT-CTC, English) | 1.1B | 2024-05-07 | cc-by-4.0 | Download | |
| parakeet-tdt_ctc-0.6b-ja | speech recognition (hybrid TDT-CTC, Japanese) | 0.6B | 2024-05-13 | cc-by-4.0 | Download | |
| parakeet-tdt_ctc-110m | speech recognition (hybrid TDT-CTC, English) | 110M | 2024-09-17 | cc-by-4.0 | Download | |
| parakeet-tdt-0.6b-v2 | speech recognition (TDT, English) | 0.6B | 2025-04-15 | cc-by-4.0 | Download | |
| parakeet-rnnt-110m-da-dk | speech recognition (RNN-T, Danish) | 110M | 2025-05-28 | nvidia-open-model-license | Download | |
| parakeet-tdt-0.6b-v3 Parakeet TDT 0.6B v3 is NVIDIA's open-weight speech recognition model for 25 European languages, a 0.6B-parameter FastConformer encoder with a TDT decoder that detects the language and returns punctuated text with word and segment timestamps, under CC BY 4.0. (The repository also holds NVIDIA's own Q8_0 GGUF and the .nemo archive; there is no separate build repository. The card says it extends parakeet-tdt-0.6b-v2 from English to 25 European languages.) | speech recognition (TDT, 25 European languages) | 0.6B (627M per the safetensors count) | 2025-08-04 | cc-by-4.0 | Our listing (B) · Download | |
| parakeet_realtime_eou_120m-v1 | streaming speech recognition with end-of-utterance detection | 120M | 2025-10-10 | nvidia-open-model-license | Download | |
| multitalker-parakeet-streaming-0.6b-v1 | streaming multi-speaker speech recognition | 0.6B | 2025-10-15 | nvidia-open-model-license | Download | |
| parakeet-ctc-0.6b-Vietnamese | speech recognition (CTC, Vietnamese) | 0.6B | 2026-01-15 | nvidia-open-model-license | Download | |
| parakeet-unified-en-0.6b | speech recognition (English) | 0.6B | 2026-04-07 | nvidia-open-model-license | Download |
Cosmos-Reason2 from 2025-12-12
NVIDIA's physical-AI reasoning vision-language models, fine-tuned from Qwen3-VL Instruct models of the same size (each card's base_model). A derivative NVIDIA publishes under its own name, so base is set on every member. Released is the Hub createdAt. Licence is the NVIDIA Open Model License (license_name nvidia-open-model-license). NVIDIA publishes no separate quantised builds of these.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Cosmos-Reason2-2B Cosmos-Reason2-2B is NVIDIA's open-weight vision language model for physical AI and robotics, a 2.4B-parameter fine-tune of Qwen3-VL-2B-Instruct that reasons over video and images. The repository is gated behind a click-through licence. | reasoning vision-language (image and video in, text out) | 2B (2.44B per the safetensors count) | 2025-12-12 | nvidia-open-model-license | Qwen/Qwen3-VL-2B-Instruct | Our listing (C) · Download (gated) |
| Cosmos-Reason2-8B | reasoning vision-language (image and video in, text out) | 8B (8.77B per the safetensors count) | 2025-12-12 | nvidia-open-model-license | Qwen/Qwen3-VL-8B-Instruct | Download (gated) |
| Cosmos-Reason2-32B | reasoning vision-language (image and video in, text out) | 32B (33.36B per the safetensors count) | 2026-04-29 | nvidia-open-model-license | Qwen/Qwen3-VL-32B-Instruct | Download |
Older versions (9 families, 43 models)
Nemotron Nano 2 from 2025-08-18
Hybrid Mamba-Transformer. NVIDIA names these NVIDIA-Nemotron-Nano-*-v2; Nemotron 3 Nano 4B was compressed from the 9B.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| NVIDIA-Nemotron-Nano-9B-v2 (Front matter also lists nvidia/NVIDIA-Nemotron-Nano-12B-v2 as a base.) | instruct (reasoning on/off) | 9B | 2025-08-18 | nvidia-open-model-license | nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base | Download · NVIDIA-Nemotron-Nano-9B-v2-FP8 · NVIDIA-Nemotron-Nano-9B-v2-NVFP4 |
| NVIDIA-Nemotron-Nano-9B-v2-Base | base | 9B | 2025-08-18 | nvidia-open-model-license | nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base | Download |
| NVIDIA-Nemotron-Nano-12B-v2 | instruct (reasoning on/off) | 12B | 2025-08-29 | nvidia-open-model-license | nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base | Download |
| NVIDIA-Nemotron-Nano-12B-v2-Base | base | 12B | 2025-08-18 | nvidia-open-model-license | Download | |
| NVIDIA-Nemotron-Nano-9B-v2-Japanese | instruct, Japanese | 9B | 2026-02-17 | nvidia-nemotron-open-model-license | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | Download |
| NVIDIA-Nemotron-Nano-12B-v2-VL | instruct, vision-language | 12B | 2025-10-28 | nvidia-open-model-license | Download · NVIDIA-Nemotron-Nano-12B-v2-VL-FP8 · NVIDIA-Nemotron-Nano-12B-v2-VL-NVFP4-QAD |
Nemotron-H from 2025-03-19
Hybrid Mamba-Transformer research models under NVIDIA's internal scientific research and development licence (non-commercial).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Nemotron-H-8B-Base-8K | base | 8B | 2025-03-19 | nvidia-internal-scientific-research-and-development-model-license | Download | |
| Nemotron-H-4B-Base-8K (Card gives 10/23/2025 as its Hugging Face date; the repository was created 2025-03-20.) | base | 4B | 2025-10-23 | nvidia-internal-scientific-research-and-development-model-license | nvidia/Nemotron-H-8B-Base-8K | Download |
| Nemotron-H-4B-Instruct-128K (Card gives 10/23/2025 as its Hugging Face date; the repository was created 2025-04-15.) | instruct | 4B | 2025-10-23 | nvidia-internal-scientific-research-and-development-model-license | nvidia/Nemotron-H-4B-Base-8K | Download |
| Nemotron-H-56B-Base-8K | base | 56B | 2025-04-08 | nvidia-internal-scientific-research-and-development-model-license | Download | |
| Nemotron-H-47B-Base-8K | base | 47B | 2025-04-08 | nvidia-internal-scientific-research-and-development-model-license | nvidia/Nemotron-H-56B-Base-8K | Download |
| Nemotron-H-8B-Reasoning-128K | reasoning | 8B | 2025-06-06 | nvidia-internal-scientific-research-and-development-model-license | nvidia/Nemotron-H-8B-Base-8K | Download · Nemotron-H-8B-Reasoning-128K-FP8 |
| Nemotron-H-47B-Reasoning-128K | reasoning | 47B | 2025-06-06 | nvidia-internal-scientific-research-and-development-model-license | nvidia/Nemotron-H-47B-Base-8K | Download · Nemotron-H-47B-Reasoning-128K-FP8 |
Llama Nemotron (Nano, Super, Ultra) from 2025-03-16
Derivatives of Meta Llama 3.1 and 3.3 (pruned with neural architecture search and post-trained for reasoning).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama-3.1-Nemotron-Nano-8B-v1 | reasoning (on/off) | 8B | 2025-03-16 | nvidia-open-model-license | meta-llama/Llama-3.1-8B-Instruct | Download |
| Llama-3_3-Nemotron-Super-49B-v1 | reasoning (on/off) | 49B | 2025-03-16 | nvidia-open-model-license | meta-llama/Llama-3.3-70B-Instruct | Download · Llama-3_3-Nemotron-Super-49B-v1-FP8 |
| Llama-3_1-Nemotron-Ultra-253B-v1 | reasoning (on/off) | 253B | 2025-04-07 | nvidia-open-model-license | meta-llama/Llama-3.1-405B-Instruct | Download · Llama-3_1-Nemotron-Ultra-253B-v1-FP8 |
| Llama-3_1-Nemotron-Ultra-253B-CPT-v1 | continued-pretraining checkpoint (before post-training) | 253B | 2025-04-08 | nvidia-open-model-license | meta-llama/Llama-3.1-405B-Instruct | Download |
| Llama-3.1-Nemotron-Nano-4B-v1.1 | reasoning (on/off) | 4B | 2025-05-03 | nvidia-open-model-license | nvidia/Llama-3.1-Minitron-4B-Width-Base | Download |
| Llama-3_3-Nemotron-Super-49B-GenRM | generative reward model | 49B | 2025-06-27 | nvidia-open-model-license | nvidia/Llama-3_3-Nemotron-Super-49B-v1 | Download |
| Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual | generative reward model, multilingual | 49B | 2025-05-28 | nvidia-open-model-license | nvidia/Llama-3_3-Nemotron-Super-49B-v1 | Download |
| Llama-3.1-Nemotron-Nano-VL-8B-V1 | instruct, vision-language (documents) | 8B | 2025-06-03 | nvidia-open-model-license | Download · Llama-3.1-Nemotron-Nano-VL-8B-V1-FP4-QAD · Llama-3.1-Nemotron-Nano-VL-8B-V1-mcore | |
| Llama-3_3-Nemotron-Super-49B-v1_5 | reasoning (on/off) | 49B | 2025-07-25 | nvidia-open-model-license | meta-llama/Llama-3.3-70B-Instruct | Download · Llama-3_3-Nemotron-Super-49B-v1_5-FP8 · Llama-3_3-Nemotron-Super-49B-v1_5-NVFP4 |
Llama 3.1 Nemotron (2024) from 2024-09-22
Fine-tuned or NAS-distilled from Llama 3.1 70B Instruct.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Llama-3_1-Nemotron-51B-Instruct | instruct | 51B | 2024-09-22 | nvidia-open-model-license | meta-llama/Llama-3.1-70B-Instruct | Download |
| Llama-3.1-Nemotron-70B-Instruct-HF | instruct | 70B | 2024-10-12 | llama3.1 | meta-llama/Llama-3.1-70B-Instruct | Download · Llama-3.1-Nemotron-70B-Instruct |
Minitron from 2024-07-19
Base models pruned and distilled from larger models.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Minitron-4B-Base | base | 4B | 2024-07-19 | nvidia-open-model-license | Nemotron-4 15B (not published on Hugging Face) | Download |
| Minitron-8B-Base | base | 8B | 2024-07-19 | nvidia-open-model-license | Nemotron-4 15B (not published on Hugging Face) | Download |
| Llama-3.1-Minitron-4B-Width-Base | base | 4B | 2024-08-13 | nvidia-open-model-license | meta-llama/Llama-3.1-8B | Download |
| Llama-3.1-Minitron-4B-Depth-Base | base | 4B | 2024-08-13 | nvidia-open-model-license | meta-llama/Llama-3.1-8B | Download |
| Mistral-NeMo-Minitron-8B-Base | base | 8B | 2024-08-19 | nvidia-open-model-license | Mistral-NeMo 12B | Download |
| Mistral-NeMo-Minitron-8B-Instruct | instruct | 8B | 2024-10-02 | nvidia-open-model-license | nvidia/Mistral-NeMo-Minitron-8B-Base | Download |
Nemotron-4 from 2024-06-14
Nemotron-4 15B itself was not published; the Mini 4B models descend from it through Minitron-4B.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Nemotron-4-340B-Base | base | 340B | 2024-06-14 | nvidia-open-model-license | Download | |
| Nemotron-4-340B-Instruct | instruct | 340B | 2024-06-13 | nvidia-open-model-license | nvidia/Nemotron-4-340B-Base | Download |
| Nemotron-4-340B-Reward | reward model | 340B | 2024-06-13 | nvidia-open-model-license | Download | |
| Nemotron-Mini-4B-Instruct | instruct | 4B | 2024-09-10 | nvidia-open-model-license | nvidia/Minitron-4B-Base | Download · Nemotron-Mini-4B-Instruct-ONNX-INT4 |
| Nemotron-4-Mini-Hindi-4B-Base | base, Hindi/English | 4B | 2024-10-22 | nvidia-open-model-license | nvidia/Minitron-4B-Base | Download |
| Nemotron-4-Mini-Hindi-4B-Instruct | instruct, Hindi/English | 4B | 2024-10-22 | nvidia-open-model-license | nvidia/Nemotron-4-Mini-Hindi-4B-Base | Download |
Nemotron-3 8B (2023) from 2023-11-14
Not the same line as Nemotron 3 (2025): an earlier 8B family that reused the name. All five repositories are gated (manual approval, login required); the cards could not be read without signing in, so licence comes from the Hub API.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| nemotron-3-8b-base-4k | base | 8B | 2023-11-14 | nv-ai-foundation-models-license | Download (gated) | |
| nemotron-3-8b-chat-4k-sft | chat (SFT) | 8B | 2023-11-15 | nv-ai-foundation-models-license | Download (gated) | |
| nemotron-3-8b-chat-4k-rlhf | chat (RLHF) | 8B | 2023-11-14 | nv-ai-foundation-models-license | Download (gated) | |
| nemotron-3-8b-chat-4k-steerlm | chat (SteerLM) | 8B | 2023-11-15 | nv-ai-foundation-models-license | Download (gated) | |
| nemotron-3-8b-qa-4k | question answering | 8B | 2023-11-15 | nv-ai-foundation-models-license | Download (gated) |
Nemotron Speech from 2025-12-17
NVIDIA's earlier cache-aware streaming ASR model under the Nemotron name, English only by its name; nemotron-3.5-asr-streaming-0.6b (in Nemotron 3) is the multilingual successor in the same streaming line. Placement is ours.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| nemotron-speech-streaming-en-0.6b | speech recognition (cache-aware streaming, English) | 0.6B | 2025-12-17 | nvidia-open-model-license | Download |
Cosmos-Reason1 from 2025-04-18
Predecessor of Cosmos-Reason2, fine-tuned from Qwen2.5-VL-7B-Instruct (card base_model). Not gated.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Cosmos-Reason1-7B | reasoning vision-language (image and video in, text out) | 7B (8.29B per the safetensors count) | 2025-04-18 | nvidia-open-model-license | Qwen/Qwen2.5-VL-7B-Instruct | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Cosmos-Reason2-2B 3, Nemotron 3.5 Lightning 30B A3B 5, Nemotron 3.5 Super VL 4, Nemotron 3 Nano 30B A3B 5, Nemotron 3 Nano 4B 5, Nemotron 3 Nano Omni 30B A3B Reasoning 5, Nemotron 3 Super 120B A12B 5, Nemotron 3 Ultra 550B A55B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
OpenAI
Publishes from huggingface.co/openai, github.com/openai. Checked 2026-10-21.
gpt-oss from 2025-08-05
OpenAI's only open-weight language-model family since GPT-2. Both base sizes ship as one MXFP4 safetensors repository each (the original/ and metal/ folders sit inside the same repo), so there are no separate publisher builds. gpt-oss-safeguard is OpenAI's own fine-tune of gpt-oss for policy-based safety classification; its repos were created on 2025-09-18 and OpenAI announced them on 2025-10-29 (blog linked from the card, not read). No base (pre-trained only) checkpoints were released. OpenAI's openai/privacy-filter (2026-04, 1.5B, token-classification PII detector built from a gpt-oss-like checkpoint) is not a text-generation model and is left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| gpt-oss-120b gpt-oss-120b is OpenAI's open-weight mixture-of-experts reasoning model, 116.8B parameters with 5.1B active and a 131,072-token context, published on Hugging Face in MXFP4 under Apache 2.0. | reasoning | 117B total, 5.1B active (MoE) | 2025-08-05 | apache-2.0 | Our listing (BB) · Download | |
| gpt-oss-20b gpt-oss-20b is OpenAI's open-weight mixture-of-experts reasoning model, 21B parameters with 3.6B active and a 131,072-token context, published on Hugging Face in MXFP4 under Apache 2.0 and trained on the harmony prompt format. | reasoning | 21B total, 3.6B active (MoE) | 2025-08-05 | apache-2.0 | Our listing (B) · Download | |
| gpt-oss-safeguard-120b | safety reasoning (policy classifier) | 117B total, 5.1B active (MoE) | 2025-10-29 | apache-2.0 | openai/gpt-oss-120b | Download |
| gpt-oss-safeguard-20b | safety reasoning (policy classifier) | 21B total, 3.6B active (MoE) | 2025-10-29 | apache-2.0 | openai/gpt-oss-20b | Download |
Older versions (3 families, 6 models)
circuit-sparsity from 2025-12-11
Research release, not a general-purpose model or a gpt-oss successor: a small weight-sparse transformer trained on Python code from Gao et al. 2025 (the card calls it csp_yolo2). The card says weights for the paper's other models are at github.com/openai/circuit_sparsity (not read: GitHub returned 403 to this session). Placed here because it is an OpenAI text-generation release outside gpt-oss.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| circuit-sparsity (csp_yolo2) | base (research) | 0.42B (safetensors total from Hub API; card states none) | 2025-12-11 | apache-2.0 | Download |
GPT-2 from 2019-02-14
Not in OpenAI's Hugging Face organisation. OpenAI's own download location is the archived GitHub repo openai/gpt-2 (Modified MIT licence), whose download_model.py fetches TensorFlow checkpoints from openaipublic.blob.core.windows.net/gpt-2/models/<size>/; all four hparams.json files answered HTTP 200 on 2026-10-21. The huggingface.co/openai-community copies (gpt2, gpt2-medium, gpt2-large, gpt2-xl) are maintained by Hugging Face's transformers team, not OpenAI, so they are not used. Sizes are the repo's corrected counts (formerly called 117M and 345M). Staged-release dates are from OpenAI's 2019 posts, recalled rather than re-read today.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GPT-2 124M | base | 124M | 2019-02-14 | Modified MIT | Download | |
| GPT-2 355M | base | 355M | 2019-05-03 | Modified MIT | Download | |
| GPT-2 774M | base | 774M | 2019-08-20 | Modified MIT | Download | |
| GPT-2 1558M (1.5B) | base | 1558M (1.5B) | 2019-11-05 | Modified MIT | Download |
GPT (GPT-1) from 2018-06-11
Archived GitHub repo openai/finetune-transformer-lm (MIT) holds the pre-trained weights as numpy files in model/. The repo README does not state a parameter count (the paper's figure is about 117M, not re-read). The huggingface.co/openai-community/openai-gpt copy is Hugging Face-maintained. Release date recalled from OpenAI's June 2018 post, not re-read.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GPT (finetune-transformer-lm) | base | not stated in repo | 2018-06-11 | mit | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: gpt-oss-120b 5, gpt-oss-20b 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
OpenBMB (MiniCPM)
Publishes from huggingface.co/openbmb. Checked 2026-10-21.
MiniCPM5 from 2026-05-21
MiniCPM5-1B came first (repositories created 2026-05-21), MiniCPM5-2B second. Dates are our listing's firstReleased for MiniCPM5-2B and otherwise the repository's createdAt from the Hub API. Every repository is also mirrored on ModelScope (OpenBMB org), per the cards. Nothing is gated.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM5-2B MiniCPM5-2B is OpenBMB's open-weight dense language model of 2.5B parameters for on-device and local use, with thinking, XML tool calls and a 131,072-token context, published on Hugging Face in BF16, GGUF, MLX and GPTQ under Apache 2.0. | instruct (thinking on/off, final release after RL) | 2B (2.5B total per the safetensors index) | 2026-09-07 | apache-2.0 | Our listing (C) · Download · MiniCPM5-2B-GGUF · MiniCPM5-2B-MLX · MiniCPM5-2B-GPTQ · MiniCPM5-2B-DSpark · MiniCPM5-2B-DSpark-GGUF | |
| MiniCPM5-2B-SFT | SFT-only checkpoint (before RL) | 2B | 2026-08-27 | apache-2.0 | Download | |
| MiniCPM5-2B-Midtrain | mid-training checkpoint (before SFT) | 2B | 2026-09-01 | apache-2.0 | Download | |
| MiniCPM5-2B-Base | base | 2B | 2026-08-27 | apache-2.0 | Download | |
| MiniCPM5-1B | instruct (final release after RL) | 1B | 2026-05-21 | apache-2.0 | Download · MiniCPM5-1B-GGUF · MiniCPM5-1B-MLX | |
| MiniCPM5-1B-SFT | SFT-only checkpoint (before RL) | 1B | 2026-05-21 | apache-2.0 | Download | |
| MiniCPM5-1B-Base | base | 1B | 2026-05-21 | apache-2.0 | Download |
MiniCPM-o 4 from 2026-02-03
OpenBMB's omni-modal line (vision, speech in and out, full-duplex live streaming, per the card's first line). OpenBMB published no MiniCPM-o 4.0; 4.5 is placed in a 'MiniCPM-o 4' family by the minor-version rule. Released is the Hub createdAt of the main repository (the GGUF and AWQ builds were created a day earlier). Not gated.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM-o 4.5 MiniCPM-o 4.5 is OpenBMB's open-weight omni-modal model of 9.4B parameters that takes images, video, audio and text and answers in text and speech, including full-duplex live conversation and voice cloning from a reference clip, under Apache 2.0. | omni (text, image, video and speech in; text and speech out) | 9.4B (9,371,787,666 per the safetensors count) | 2026-02-03 | apache-2.0 | Our listing (C) · Download · MiniCPM-o-4_5-gguf · MiniCPM-o-4_5-awq · MiniCPM-o-4_5-GPTQ |
Older versions (5 families, 22 models)
MiniCPM4 (incl. MiniCPM4.1) from 2025-06-05
BitCPM-CANN (2026-05) and MiniCPM-SALA (2026-02) are placed here because their cards tie them to MiniCPM4 (BitCPM-CANN is evaluated against and, for 8B, built from MiniCPM4; MiniCPM-SALA links the MiniCPM4 technical report). Eagle and FRSpec repositories are speculative-decoding heads.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM4-8B | instruct | 8B | 2025-06-05 | apache-2.0 | Download · MiniCPM4-8B-GGUF · MiniCPM4-8B-mlx · MiniCPM4-8B-marlin-vLLM · MiniCPM4-8B-marlin-cpmcu · MiniCPM4-8B-Eagle-FRSpec · MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu · MiniCPM4-8B-Eagle-vLLM · MiniCPM4-8B-marlin-Eagle-vLLM | |
| MiniCPM4-0.5B | instruct | 0.5B | 2025-06-05 | apache-2.0 | Download · MiniCPM4-0.5B-mlx · MiniCPM4-0.5B-QAT-Int4-unquantized · MiniCPM4-0.5B-QAT-Int4-GPTQ-format · MiniCPM4-0.5B-QAT-Int4-GGUF | |
| BitCPM4-0.5B (Base named on the MiniCPM4-8B card.) | ternary (1.58-bit) quantisation-aware trained | 0.5B | 2025-06-05 | apache-2.0 | openbmb/MiniCPM4-0.5B | Download · BitCPM4-0.5B-GGUF |
| BitCPM4-1B (The MiniCPM4-8B card says it was made from "MiniCPM3-1B", which is not a published repository, so no base is given.) | ternary (1.58-bit) quantisation-aware trained | 1B | 2025-06-05 | apache-2.0 | Download · BitCPM4-1B-GGUF | |
| MiniCPM4-Survey | agent fine-tune (survey writing) | 8B | 2025-06-05 | apache-2.0 | openbmb/MiniCPM4-8B | Download |
| MiniCPM4-MCP | agent fine-tune (MCP tool calling) | 8B | 2025-06-05 | apache-2.0 | openbmb/MiniCPM4-8B | Download |
| MiniCPM4.1-8B | instruct (hybrid reasoning) | 8B | 2025-09-02 | apache-2.0 | Download · MiniCPM4.1-8B-GGUF · MiniCPM4.1-8B-MLX · MiniCPM4.1-8B-GPTQ · MiniCPM4.1-8B-AutoAWQ · MiniCPM4.1-8B-Marlin · MiniCPM4.1-8B-Eagle3 | |
| MiniCPM-SALA (Card says it was made by Transformer-to-hybrid continued training of pre-trained weights but does not name the model.) | long-context hybrid sparse/linear attention | 9B (9.5B per the safetensors index) | 2026-02-11 | apache-2.0 | Download | |
| BitCPM-CANN-8B | ternary (1.58-bit) QAT, pseudo-quantised | 8B | 2026-05-15 | apache-2.0 | openbmb/MiniCPM4-8B | Download · BitCPM-CANN-8B-gguf · BitCPM-CANN-8B-unquantized |
| BitCPM-CANN-3B | ternary (1.58-bit) QAT, pseudo-quantised | 3B | 2026-05-15 | apache-2.0 | Download · BitCPM-CANN-3B-gguf · BitCPM-CANN-3B-unquantized | |
| BitCPM-CANN-1B | ternary (1.58-bit) QAT, pseudo-quantised | 1B | 2026-05-15 | apache-2.0 | Download · BitCPM-CANN-1B-gguf · BitCPM-CANN-1B-unquantized | |
| BitCPM-CANN-0.5B | ternary (1.58-bit) QAT, pseudo-quantised | 0.5B | 2026-05-15 | apache-2.0 | Download · BitCPM-CANN-0.5B-gguf · BitCPM-CANN-0.5B-unquantized |
MiniCPM3 from 2024-09-03
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM3-4B | instruct (function calling) | 4B | 2024-09-03 | apache-2.0 | Download · MiniCPM3-4B-GPTQ-Int4 · MiniCPM3-4B-GGUF | |
| MiniCPM3-RAG-LoRA | RAG LoRA adapter | 4B (LoRA adapter) | 2024-09-04 | not stated on card or front matter | openbmb/MiniCPM3-4B | Download |
MiniCPM 2.0 from 2024-04-07
Grouped as OpenBMB's April 2024 MiniCPM 2.0 release (the MiniCPM3 card refers to MiniCPM1.0/MiniCPM2.0). MiniCPM-S-1B (June 2024) is a ProSparse variant of MiniCPM-1B and sits here too.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM-2B-128k | instruct, 128K context | 2B (2.4B non-embedding per the card) | 2024-04-09 | not stated on card or front matter | openbmb/MiniCPM-2B-sft-bf16 | Download |
| MiniCPM-MoE-8x2B | instruct (SFT, no RLHF) | 8x2B | 2024-04-07 | not stated on card or front matter | Download | |
| MiniCPM-1B-sft-bf16 | instruct (SFT) | 1B (1.2B non-embedding per the card) | 2024-04-10 | General Model License (GML), per the card: free for academic research, commercial use after registration with OpenBMB/ModelBest | Download | |
| MiniCPM-S-1B-sft | activation-sparse (ProSparse) instruct | 1B | 2024-06-03 | apache-2.0 | openbmb/MiniCPM-1B-sft-bf16 | Download · MiniCPM-S-1B-sft-llama-format · MiniCPM-S-1B-sft-gguf |
MiniCPM 1.0 from 2024-01-29
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM-2B-sft-bf16 | instruct (SFT) | 2B (2.4B non-embedding per the card) | 2024-01-29 | General Model License (GML), per the card: free for academic research, commercial use after registration with OpenBMB/ModelBest | Download · MiniCPM-2B-sft-fp32 · MiniCPM-2B-sft-int4 · MiniCPM-2B-sft-bf16-llama-format · MiniCPM-2B-sft-fp32-llama-format | |
| MiniCPM-2B-dpo-bf16 | instruct (DPO) | 2B (2.4B non-embedding per the card) | 2024-01-30 | General Model License (GML), per the card: free for academic research, commercial use after registration with OpenBMB/ModelBest | Download · MiniCPM-2B-dpo-fp32 · MiniCPM-2B-dpo-fp16 · MiniCPM-2B-dpo-int4 · MiniCPM-2B-dpo-bf16-llama-format · minicpm-dpo-bf16-ggml-model-q4_0 | |
| MiniCPM-2B-history | intermediate training checkpoints (research) | 2B | 2024-03-08 | General Model License (GML), per the card: free for academic research, commercial use after registration with OpenBMB/ModelBest | Download |
MiniCPM-o 2 (MiniCPM-o 2.6) from 2025-01-12
The first MiniCPM-o release, earlier version of the omni line in families (not of the MiniCPM language models above). Released is the Hub createdAt.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiniCPM-o 2.6 | omni (text, image, video and speech in; text and speech out) | 8.7B (8,674,997,028 per the safetensors count) | 2025-01-12 | apache-2.0 | Download · MiniCPM-o-2_6-gguf · MiniCPM-o-2_6-int4 · MiniCPM-o-2_6-GPTQ |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: MiniCPM5-2B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
OpenGVLab (Shanghai AI Laboratory)
Publishes from huggingface.co/OpenGVLab. Checked 2026-10-21.
InternVL3 (3.0, 3.5) from 2025-04-10
InternVL3 (April 2025) and InternVL3.5 (25 August 2025, Flash editions 28 September 2025). Every model pairs OpenGVLab's InternViT vision encoder with another publisher's language model, declared in the Pretrained repositories' metadata as a merge: Qwen2.5 and InternLM3 for InternVL3, Qwen3 and gpt-oss-20b for InternVL3.5. base is that language model. Editions are Pretrained, Instruct (SFT), MPO and the final model; the -hf and -HF repositories (Transformers format) and AWQ repositories are builds. The listed InternVL3_5-GPT-OSS-20B-A4B-Preview pairs InternViT-300M-448px-V2_5 with openai/gpt-oss-20b; our listing points at its -HF build, which holds the full 21.23B weights, while the original-format repository shows only 392M on the Hub. Apache 2.0 except InternVL3-9B (MIT) and the 78B models (Qwen licence). Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| InternVL3_5-GPT-OSS-20B-A4B-Preview InternVL3.5-GPT-OSS-20B-A4B-Preview is OpenGVLab's open-weight vision-language preview that joins OpenAI's gpt-oss-20b mixture-of-experts model with the InternViT-300M vision encoder, 21.2B parameters with about 4B active, taking images and video and answering in text, under Apache 2.0. | instruct (vision-language; preview) | 20B total, 4B active (MoE) per the name (Hub safetensors count 392M for this repository; the -HF build counts 21.23B) | 2025-08-25 | apache-2.0 | openai/gpt-oss-20b | Our listing (D) · Download · InternVL3_5-GPT-OSS-20B-A4B-Preview-HF |
| InternVL3_5-8B-Flash | instruct (vision-language) | 8B (Hub safetensors count 8.81B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-8B | Download |
| InternVL3_5-4B-Flash | instruct (vision-language) | 4B (Hub safetensors count 4.91B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-4B | Download |
| InternVL3_5-38B-Flash | instruct (vision-language) | 38B (Hub safetensors count 39.56B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-32B | Download |
| InternVL3_5-30B-A3B-Flash | instruct (vision-language) | 30B total, 3B active (MoE) (Hub safetensors count 30.99B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-30B-A3B | Download |
| InternVL3_5-2B-Flash | instruct (vision-language) | 2B (Hub safetensors count 2.49B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-1.7B | Download |
| InternVL3_5-241B-A28B-Flash | instruct (vision-language) | 241B total, 28B active (MoE) (Hub safetensors count 241.71B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-235B-A22B | Download |
| InternVL3_5-1B-Flash | instruct (vision-language) | 1B (Hub safetensors count 1.15B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| InternVL3_5-14B-Flash | instruct (vision-language) | 14B (Hub safetensors count 15.50B) | 2025-09-28 | apache-2.0 | Qwen/Qwen3-14B | Download |
| InternVL3_5-8B-Pretrained | base | 8B (Hub safetensors count 8.53B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-8B | Download |
| InternVL3_5-8B-MPO | instruct (MPO stage) | 8B (Hub safetensors count 8.53B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-8B | Download |
| InternVL3_5-8B-Instruct | instruct (vision-language) | 8B (Hub safetensors count 8.53B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-8B | Download |
| InternVL3_5-8B | instruct (vision-language) | 8B (Hub safetensors count 8.53B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-8B | Download · InternVL3_5-8B-HF |
| InternVL3_5-4B-Pretrained | base | 4B (Hub safetensors count 4.73B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-4B | Download |
| InternVL3_5-4B-MPO | instruct (MPO stage) | 4B (Hub safetensors count 4.73B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-4B | Download |
| InternVL3_5-4B-Instruct | instruct (vision-language) | 4B (Hub safetensors count 4.73B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-4B | Download |
| InternVL3_5-4B | instruct (vision-language) | 4B (Hub safetensors count 4.73B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-4B | Download · InternVL3_5-4B-HF |
| InternVL3_5-38B-Pretrained | base | 38B (Hub safetensors count 38.39B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-32B | Download |
| InternVL3_5-38B-MPO | instruct (MPO stage) | 38B (Hub safetensors count 38.39B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-32B | Download |
| InternVL3_5-38B-Instruct | instruct (vision-language) | 38B (Hub safetensors count 38.39B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-32B | Download |
| InternVL3_5-38B | instruct (vision-language) | 38B (Hub safetensors count 38.39B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-32B | Download · InternVL3_5-38B-HF |
| InternVL3_5-30B-A3B-Pretrained | base | 30B total, 3B active (MoE) (Hub safetensors count 30.85B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-30B-A3B | Download |
| InternVL3_5-30B-A3B-MPO | instruct (MPO stage) | 30B total, 3B active (MoE) (Hub safetensors count 30.85B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-30B-A3B | Download |
| InternVL3_5-30B-A3B-Instruct | instruct (vision-language) | 30B total, 3B active (MoE) (Hub safetensors count 30.85B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-30B-A3B | Download |
| InternVL3_5-30B-A3B | instruct (vision-language) | 30B total, 3B active (MoE) (Hub safetensors count 30.85B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-30B-A3B | Download · InternVL3_5-30B-A3B-HF |
| InternVL3_5-2B-Pretrained | base | 2B (Hub safetensors count 2.35B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-1.7B | Download |
| InternVL3_5-2B-MPO | instruct (MPO stage) | 2B (Hub safetensors count 2.35B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-1.7B | Download |
| InternVL3_5-2B-Instruct | instruct (vision-language) | 2B (Hub safetensors count 2.35B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-1.7B | Download |
| InternVL3_5-2B | instruct (vision-language) | 2B (Hub safetensors count 2.35B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-1.7B | Download · InternVL3_5-2B-HF |
| InternVL3_5-241B-A28B-Pretrained | base | 241B total, 28B active (MoE) (Hub safetensors count 240.70B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-235B-A22B | Download |
| InternVL3_5-241B-A28B-MPO | instruct (MPO stage) | 241B total, 28B active (MoE) (Hub safetensors count 240.70B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-235B-A22B | Download |
| InternVL3_5-241B-A28B-Instruct | instruct (vision-language) | 241B total, 28B active (MoE) (Hub safetensors count 240.70B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-235B-A22B | Download |
| InternVL3_5-241B-A28B | instruct (vision-language) | 241B total, 28B active (MoE) (Hub safetensors count 240.70B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-235B-A22B | Download · InternVL3_5-241B-A28B-HF |
| InternVL3_5-1B-Pretrained | base | 1B (Hub safetensors count 1.06B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| InternVL3_5-1B-MPO | instruct (MPO stage) | 1B (Hub safetensors count 1.06B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| InternVL3_5-1B-Instruct | instruct (vision-language) | 1B (Hub safetensors count 1.06B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-0.6B | Download |
| InternVL3_5-1B | instruct (vision-language) | 1B (Hub safetensors count 1.06B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-0.6B | Download · InternVL3_5-1B-HF |
| InternVL3_5-14B-Pretrained | base | 14B (Hub safetensors count 15.12B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-14B | Download |
| InternVL3_5-14B-MPO | instruct (MPO stage) | 14B (Hub safetensors count 15.12B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-14B | Download |
| InternVL3_5-14B-Instruct | instruct (vision-language) | 14B (Hub safetensors count 15.12B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-14B | Download |
| InternVL3_5-14B | instruct (vision-language) | 14B (Hub safetensors count 15.12B) | 2025-08-25 | apache-2.0 | Qwen/Qwen3-14B | Download · InternVL3_5-14B-HF |
| InternVL3-9B-Pretrained | base | 9B (Hub safetensors count 9.14B) | 2025-04-17 | mit | internlm/internlm3-8b-instruct | Download |
| InternVL3-8B-Pretrained | base | 8B (Hub safetensors count 7.94B) | 2025-04-17 | apache-2.0 | Qwen/Qwen2.5-7B | Download |
| InternVL3-78B-Pretrained | base | 78B (Hub safetensors count 78.41B) | 2025-04-17 | other (qwen) | Qwen/Qwen2.5-72B | Download |
| InternVL3-38B-Pretrained | base | 38B (Hub safetensors count 38.39B) | 2025-04-17 | apache-2.0 | Qwen/Qwen2.5-32B | Download |
| InternVL3-2B-Pretrained | base | 2B (Hub safetensors count 2.09B) | 2025-04-17 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download |
| InternVL3-1B-Pretrained | base | 1B (Hub safetensors count 938M) | 2025-04-17 | apache-2.0 | Qwen/Qwen2.5-0.5B | Download |
| InternVL3-14B-Pretrained | base | 14B (Hub safetensors count 15.12B) | 2025-04-17 | apache-2.0 | Qwen/Qwen2.5-14B | Download |
| InternVL3-9B-Instruct | instruct (vision-language) | 9B (Hub safetensors count 9.14B) | 2025-04-16 | mit | internlm/internlm3-8b-instruct | Download |
| InternVL3-8B-Instruct | instruct (vision-language) | 8B (Hub safetensors count 7.94B) | 2025-04-16 | apache-2.0 | Qwen/Qwen2.5-7B | Download |
| InternVL3-78B-Instruct | instruct (vision-language) | 78B (Hub safetensors count 78.41B) | 2025-04-16 | other (qwen) | Qwen/Qwen2.5-72B | Download |
| InternVL3-38B-Instruct | instruct (vision-language) | 38B (Hub safetensors count 38.39B) | 2025-04-16 | apache-2.0 | Qwen/Qwen2.5-32B | Download |
| InternVL3-2B-Instruct | instruct (vision-language) | 2B (Hub safetensors count 2.09B) | 2025-04-16 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download |
| InternVL3-1B-Instruct | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2025-04-16 | apache-2.0 | Qwen/Qwen2.5-0.5B | Download |
| InternVL3-14B-Instruct | instruct (vision-language) | 14B (Hub safetensors count 15.12B) | 2025-04-16 | apache-2.0 | Qwen/Qwen2.5-14B | Download |
| InternVL3-9B | instruct (vision-language) | 9B (Hub safetensors count 9.14B) | 2025-04-10 | mit | internlm/internlm3-8b-instruct | Download · InternVL3-9B-AWQ |
| InternVL3-8B | instruct (vision-language) | 8B (Hub safetensors count 7.94B) | 2025-04-10 | apache-2.0 | Qwen/Qwen2.5-7B | Download · InternVL3-8B-AWQ · InternVL3-8B-hf |
| InternVL3-78B | instruct (vision-language) | 78B (Hub safetensors count 78.41B) | 2025-04-10 | other (qwen) | Qwen/Qwen2.5-72B | Download · InternVL3-78B-AWQ · InternVL3-78B-hf |
| InternVL3-38B | instruct (vision-language) | 38B (Hub safetensors count 38.39B) | 2025-04-10 | apache-2.0 | Qwen/Qwen2.5-32B | Download · InternVL3-38B-AWQ · InternVL3-38B-hf |
| InternVL3-2B | instruct (vision-language) | 2B (Hub safetensors count 2.09B) | 2025-04-10 | apache-2.0 | Qwen/Qwen2.5-1.5B | Download · InternVL3-2B-AWQ · InternVL3-2B-hf |
| InternVL3-1B | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2025-04-10 | apache-2.0 | Qwen/Qwen2.5-0.5B | Download · InternVL3-1B-AWQ · InternVL3-1B-hf |
| InternVL3-14B | instruct (vision-language) | 14B (Hub safetensors count 15.12B) | 2025-04-10 | apache-2.0 | Qwen/Qwen2.5-14B | Download · InternVL3-14B-AWQ · InternVL3-14B-hf |
Older versions (3 families, 41 models)
InternVL2.5 from 2024-11-20
Earlier InternVL generations, newest first, all built on other publishers' language models (base as declared).
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| InternVL2_5-8B-MPO | instruct (MPO stage) | 8B (Hub safetensors count 8.08B) | 2024-12-20 | mit | internlm/internlm2_5-7b-chat | Download · InternVL2_5-8B-MPO-AWQ · InternVL2_5-8B-MPO-hf |
| InternVL2_5-78B-MPO | instruct (MPO stage) | 78B (Hub safetensors count 78.41B) | 2024-12-20 | other (qwen) | Qwen/Qwen2.5-72B-Instruct | Download · InternVL2_5-78B-MPO-AWQ |
| InternVL2_5-4B-MPO | instruct (MPO stage) | 4B (Hub safetensors count 3.71B) | 2024-12-20 | mit | Qwen/Qwen2.5-3B-Instruct | Download · InternVL2_5-4B-MPO-AWQ |
| InternVL2_5-38B-MPO | instruct (MPO stage) | 38B (Hub safetensors count 38.39B) | 2024-12-20 | mit | Qwen/Qwen2.5-32B-Instruct | Download · InternVL2_5-38B-MPO-AWQ |
| InternVL2_5-2B-MPO | instruct (MPO stage) | 2B (Hub safetensors count 2.21B) | 2024-12-20 | mit | internlm/internlm2_5-1_8b-chat | Download · InternVL2_5-2B-MPO-hf |
| InternVL2_5-26B-MPO | instruct (MPO stage) | 26B (Hub safetensors count 25.51B) | 2024-12-20 | mit | internlm/internlm2_5-20b-chat | Download · InternVL2_5-26B-MPO-AWQ |
| InternVL2_5-1B-MPO | instruct (MPO stage) | 1B (Hub safetensors count 938M) | 2024-12-20 | mit | Qwen/Qwen2.5-0.5B-Instruct | Download |
| InternVL2_5-78B | instruct (vision-language) | 78B (Hub safetensors count 78.41B) | 2024-12-02 | other (qwen) | Qwen/Qwen2.5-72B-Instruct | Download · InternVL2_5-78B-AWQ |
| InternVL2_5-8B | instruct (vision-language) | 8B (Hub safetensors count 8.08B) | 2024-11-20 | mit | internlm/internlm2_5-7b-chat | Download · InternVL2_5-8B-AWQ |
| InternVL2_5-4B | instruct (vision-language) | 4B (Hub safetensors count 3.71B) | 2024-11-20 | mit | Qwen/Qwen2.5-3B-Instruct | Download · InternVL2_5-4B-AWQ |
| InternVL2_5-38B | instruct (vision-language) | 38B (Hub safetensors count 38.39B) | 2024-11-20 | mit | Qwen/Qwen2.5-32B-Instruct | Download · InternVL2_5-38B-AWQ |
| InternVL2_5-2B | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-11-20 | mit | internlm/internlm2_5-1_8b-chat | Download |
| InternVL2_5-26B | instruct (vision-language) | 26B (Hub safetensors count 25.51B) | 2024-11-20 | mit | internlm/internlm2_5-20b-chat | Download · InternVL2_5-26B-AWQ |
| InternVL2_5-1B | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2024-11-20 | mit | Qwen/Qwen2.5-0.5B-Instruct | Download |
InternVL2 (with Mini-InternVL2 domain adaptations) from 2024-06-27
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mini-InternVL2-4B-DA-Medical | instruct (vision-language) | 4B (Hub safetensors count 4.15B) | 2024-12-07 | mit | microsoft/Phi-3-mini-128k-instruct | Download |
| Mini-InternVL2-4B-DA-DriveLM | instruct (vision-language) | 4B (Hub safetensors count 4.15B) | 2024-12-07 | mit | microsoft/Phi-3-mini-128k-instruct | Download |
| Mini-InternVL2-4B-DA-BDD | instruct (vision-language) | 4B (Hub safetensors count 4.15B) | 2024-12-07 | mit | microsoft/Phi-3-mini-128k-instruct | Download |
| Mini-InternVL2-2B-DA-Medical | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-12-07 | mit | internlm/internlm2-chat-1_8b | Download |
| Mini-InternVL2-2B-DA-DriveLM | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-12-07 | mit | internlm/internlm2-chat-1_8b | Download |
| Mini-InternVL2-2B-DA-BDD | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-12-07 | mit | internlm/internlm2-chat-1_8b | Download |
| Mini-InternVL2-1B-DA-Medical | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2024-12-07 | mit | Qwen/Qwen2-0.5B-Instruct | Download |
| Mini-InternVL2-1B-DA-DriveLM | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2024-12-07 | mit | Qwen/Qwen2-0.5B-Instruct | Download |
| Mini-InternVL2-1B-DA-BDD | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2024-12-07 | mit | Qwen/Qwen2-0.5B-Instruct | Download |
| InternVL2-8B-MPO | instruct (MPO stage) | 8B (Hub safetensors count 8.08B) | 2024-11-07 | mit | internlm/internlm2_5-7b-chat | Download |
| InternVL2-Llama3-76B | instruct (vision-language) | 76B (Hub safetensors count 76.26B) | 2024-07-15 | llama3 | NousResearch/Hermes-2-Theta-Llama-3-70B | Download · InternVL2-Llama3-76B-AWQ |
| InternVL2-1B | instruct (vision-language) | 1B (Hub safetensors count 938M) | 2024-07-08 | mit | Qwen/Qwen2-0.5B-Instruct | Download |
| InternVL2-8B | instruct (vision-language) | 8B (Hub safetensors count 8.08B) | 2024-06-27 | mit | internlm/internlm2_5-7b-chat | Download · InternVL2-8B-AWQ |
| InternVL2-4B | instruct (vision-language) | 4B (Hub safetensors count 4.15B) | 2024-06-27 | mit | microsoft/Phi-3-mini-128k-instruct | Download |
| InternVL2-40B | instruct (vision-language) | 40B (Hub safetensors count 40.07B) | 2024-06-27 | mit | NousResearch/Nous-Hermes-2-Yi-34B | Download · InternVL2-40B-AWQ |
| InternVL2-2B | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-06-27 | mit | internlm/internlm2-chat-1_8b | Download · InternVL2-2B-AWQ |
| InternVL2-26B | instruct (vision-language) | 26B (Hub safetensors count 25.51B) | 2024-06-27 | mit | internlm/internlm2-chat-20b | Download · InternVL2-26B-AWQ |
InternVL 1.x (InternVL-Chat V1.1 to V1.5, Mini-InternVL-Chat) from 2023-12-22
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Mini-InternVL-Chat-4B-V1-5 | instruct (vision-language) | 4B (Hub safetensors count 4.15B) | 2024-05-28 | mit | microsoft/Phi-3-mini-128k-instruct | Download |
| Mini-InternVL-Chat-2B-V1-5 | instruct (vision-language) | 2B (Hub safetensors count 2.21B) | 2024-05-13 | mit | internlm/internlm2-chat-1_8b | Download |
| InternVL-Chat-V1-5 | instruct (vision-language) | 25.51B (Hub safetensors count; the name states none) | 2024-04-18 | mit | internlm/internlm2-chat-20b | Download · InternVL-Chat-V1-5-AWQ |
| InternVL-Chat-V1-2-Plus | instruct (vision-language) | 40.07B (Hub safetensors count; the name states none) | 2024-02-21 | mit | NousResearch/Nous-Hermes-2-Yi-34B | Download |
| InternVL-Chat-V1-2 | instruct (vision-language) | 40.07B (Hub safetensors count; the name states none) | 2024-02-11 | mit | NousResearch/Nous-Hermes-2-Yi-34B | Download |
| InternVL-Chat-ViT-6B-Vicuna-13B-448px | instruct (vision-language) | not stated (InternViT-6B with Vicuna-13B, per the name) | 2024-02-01 | none in the card metadata | lmsys/vicuna-13b-v1.5 | Download |
| InternVL-Chat-V1-1 | instruct (vision-language) | not stated (no safetensors count on the Hub) | 2024-01-24 | mit | meta-llama/Llama-2-13b-hf | Download |
| InternVL-Chat-ViT-6B-Vicuna-7B | instruct (vision-language) | not stated (InternViT-6B with Vicuna-7B, per the name) | 2023-12-25 | none in the card metadata | lmsys/vicuna-7b-v1.5 | Download |
| InternVL-Chat-ViT-6B-Vicuna-13B | instruct (vision-language) | not stated (InternViT-6B with Vicuna-13B, per the name) | 2023-12-25 | none in the card metadata | lmsys/vicuna-13b-v1.5 | Download |
| InternVL-14B-224px | embedding (vision-language contrastive) | 14B (Hub safetensors count 13.83B) | 2023-12-22 | mit | Download |
Ornith
Publishes from huggingface.co/ornith-ai. Checked 2026-10-21.
Ornith-1 from 2026-06-21
Ornith-1.0 (June 2026) and Ornith-1.5 (August 2026) put in one major family by FAMILIES.md's rule (decimal = minor version), although Ornith's 1.5 card calls it 'a major step'. Every repository is MIT in its card metadata except the 9B and 35B-A3B MLX builds, which carry none. Ornith-1.0 cards link LICENSE files under deepreinforce-ai, the organisation's former name (it redirects to ornith-ai). The 1.0 card also names a 31B dense size that isn't published. Member dates are the Hub API's createdAt for the repository unless a listing records the public release; for gated repositories createdAt can precede the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Ornith-1.0-9B | reasoning (agentic coding) | 9B (dense) | 2026-06-21 | mit | Qwen3.5 (card: 'post-trained on top of Gemma 4 and Qwen 3.5'; config architecture is Qwen3.5) | Download · Ornith-1.0-9B-GGUF |
| Ornith-1.0-35B | reasoning (agentic coding) | 35B (MoE) | 2026-06-21 | mit | Qwen3.5 (card: 'post-trained on top of Gemma 4 and Qwen 3.5'; config architecture is Qwen3.5) | Download · Ornith-1.0-35B-FP8 · Ornith-1.0-35B-GGUF |
| Ornith-1.0-397B | reasoning (agentic coding) | 397B (MoE) | 2026-06-23 | mit | Qwen3.5 (card: 'post-trained on top of Gemma 4 and Qwen 3.5'; config architecture is Qwen3.5) | Download · Ornith-1.0-397B-FP8 |
| Ornith-1.5-9B Ornith-1.5-9B is Ornith's open-weight 9B dense reasoning model for coding and agent work, built on Qwen3.5's architecture with a 262,144-token context, published on Hugging Face in BF16, GGUF, MLX and NVFP4 and marked MIT. (Card names an Ornith-1.5-9B-Mobile build that isn't published (401).) | reasoning (agentic coding) | 9B (dense) | 2026-08-19 | mit | Qwen3.5 (card: Ornith-1.0 'developed on top of Qwen3.5 and Gemma4'; config architecture is Qwen3.5) | Our listing (C) · Download · Ornith-1.5-9B-GGUF · Ornith-1.5-9B-NVFP4 · Ornith-1.5-9B-MLX · Ornith-1.5-9B-MLX-4bit · Ornith-1.5-9B-MLX-6bit · Ornith-1.5-9B-MLX-8bit |
| Ornith-1.5-9B-DFlash | DFlash draft model for speculative decoding | 1.29B (Hub safetensors count) | 2026-09-20 | mit | ornith-ai/Ornith-1.5-9B | Download |
| Ornith-1.5-35B-A3B Ornith-1.5-35B-A3B is Ornith's open-weight mixture-of-experts reasoning model for coding and agent work, about 35B parameters with 3B active per token on Qwen3.5's architecture, with a 262,144-token context, published in BF16, FP8, NVFP4, GGUF and MLX and marked MIT. | reasoning (agentic coding) | 35B total, 3B active (MoE) | 2026-08-18 | mit | Qwen3.5 (card: Ornith-1.0 'developed on top of Qwen3.5 and Gemma4'; config architecture is Qwen3.5) | Our listing (C) · Download · Ornith-1.5-35B-A3B-FP8 · Ornith-1.5-35B-A3B-NVFP4 · Ornith-1.5-35B-A3B-GGUF · Ornith-1.5-35B-A3B-MLX · Ornith-1.5-35B-A3B-MLX-4bit · Ornith-1.5-35B-A3B-MLX-6bit · Ornith-1.5-35B-A3B-MLX-8bit |
| Ornith-1.5-35B-A3B-DFlash | DFlash draft model for speculative decoding | 0.39B (Hub safetensors count) | 2026-09-20 | mit | ornith-ai/Ornith-1.5-35B-A3B | Download |
| Ornith-1.5-122B (Released 16 Oct 2026 (createdAt); builds added 20 Oct. Card names an unpublished -Mobile build.) | reasoning (agentic coding) | 122B (MoE) | 2026-10-16 | mit | Qwen3.5 (config architecture Qwen3.5 MoE; this card's first section doesn't repeat the Qwen3.5/Gemma4 line) | Download · Ornith-1.5-122B-FP8 · Ornith-1.5-122B-NVFP4 · Ornith-1.5-122B-GGUF · Ornith-1.5-122B-MLX · Ornith-1.5-122B-MLX-4bit · Ornith-1.5-122B-MLX-8bit |
| Ornith-1.5-397B Ornith-1.5-397B is Ornith's flagship open-weight mixture-of-experts reasoning model for coding and agent work, about 397B parameters on Qwen3.5's architecture with a 262,144-token context, published on Hugging Face in BF16, FP8, NVFP4 and GGUF and marked MIT. | reasoning (agentic coding) | 397B (MoE) | 2026-08-18 | mit | Qwen3.5 (card: Ornith-1.0 'developed on top of Qwen3.5 and Gemma4'; config architecture is Qwen3.5) | Our listing (C) · Download · Ornith-1.5-397B-FP8 · Ornith-1.5-397B-NVFP4 · Ornith-1.5-397B-GGUF |
| Ornith-1.5-397B-DFlash | DFlash draft model for speculative decoding | 1.29B (Hub safetensors count) | 2026-09-20 | mit | ornith-ai/Ornith-1.5-397B | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Ornith-1.5-35B-A3B 5, Ornith-1.5-397B 4, Ornith-1.5-9B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Poolside
Publishes from huggingface.co/poolside. Checked 2026-10-21.
Laguna 2 (XS.2, XS 2.1, S 2.1) from 2026-04-28
Poolside names Laguna models by size class and generation (XS.2, XS-2.1, S-2.1). Grouped here as generation 2, with 2.1 as its minor version; the naming is our reading, Poolside publishes no family tree. XS.2 is Apache 2.0; the 2.1 models moved to OpenMDW-1.1. Poolside's FP8, NVFP4, INT4, GGUF and NVFP4-mlx repositories are builds. Its DFlash speculative-decoding drafters (about 0.46B to 1.11B) and two test fixtures are left out. Release dates for XS.2 (28 April 2026), XS 2.1 (2 July 2026) and S 2.1 (21 July 2026) are from Poolside's posts as recorded in the batch research notes; base repositories use createdAt. Laguna-S-2.1-base shows no safetensors count and no downloads on the Hub.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Laguna-S-2.1 Laguna S 2.1 is Poolside's open-weight Mixture-of-Experts model for agentic coding, 118B total and 8B active parameters, with switchable reasoning and a 1,048,576-token context, published on Hugging Face under OpenMDW-1.1. | instruct | 117.56B (Hub safetensors count; the name gives only the size class S) | 2026-07-21 | openmdw-1.1 | Our listing (C) · Download · Laguna-S-2.1-FP8 · Laguna-S-2.1-NVFP4 · Laguna-S-2.1-INT4 · Laguna-S-2.1-GGUF · Laguna-S-2.1-NVFP4-mlx | |
| Laguna-XS-2.1 Laguna XS 2.1 is Poolside's open-weight Mixture-of-Experts coding model, 33B total and 3B active parameters, with switchable reasoning and a 262,144-token context, published on Hugging Face under OpenMDW-1.1. | instruct | 33.44B (Hub safetensors count; the name gives only the size class XS) | 2026-07-02 | openmdw-1.1 | Our listing (C) · Download · Laguna-XS-2.1-FP8 · Laguna-XS-2.1-NVFP4 · Laguna-XS-2.1-INT4 · Laguna-XS-2.1-GGUF · Laguna-XS-2.1-NVFP4-mlx | |
| Laguna-S-2.1-base | base | not stated (no safetensors count on the Hub; the instruct model is 117.56B) | 2026-07-20 | openmdw-1.1 | Download | |
| Laguna-XS.2 | instruct | 33.44B (Hub safetensors count; the name gives only the size class XS) | 2026-04-28 | apache-2.0 | Download · Laguna-XS.2-FP8 · Laguna-XS.2-NVFP4 · Laguna-XS.2-INT4 |
Laguna 1 (M.1) from 2026-04-28
Laguna's first generation, one size (M). Released with XS.2 on 28 April 2026 per Poolside's post (batch research notes); the Hub repository for Laguna-M.1 itself was created 2026-06-15, after its FP8 build (2026-04-24), so it appears to have been re-created. Trained from scratch per Poolside's technical report (as recorded in the notes). No GGUF from Poolside.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Laguna-M.1 Laguna M.1 is Poolside's first-generation open-weight Mixture-of-Experts model for agentic coding, 225B total and 23B active parameters, with switchable reasoning and a 262,144-token context, published on Hugging Face under Apache 2.0. | instruct | 225.80B (Hub safetensors count; the name gives only the size class M) | 2026-04-28 | apache-2.0 | Our listing (C) · Download · Laguna-M.1-FP8 · Laguna-M.1-NVFP4 | |
| Laguna-M.1-base | base | 225.80B (Hub safetensors count; the name gives only the size class M) | 2026-06-05 | apache-2.0 | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Laguna S 2.1 5, Laguna XS 2.1 1. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
SparkLLM (XHToken)
Publishes from huggingface.co/XHToken. Checked 2026-10-21.
Spark-X2 (X2.5, X2.6) from 2026-08-24
All of XHToken's open weights: Spark-X2.5 (1.7B and 4B, 24 to 28 August 2026) and Spark-X2.6 (4B, 19 and 20 October 2026), each with a base checkpoint and XHToken's own GGUF, FP8, INT8 and (X2.6) INT4 builds. X2.6 is treated as a minor version of X2. The quantised builds' cards tag the -Base repository as what they quantise, though their names (and the GGUF file sizes) match the instruct model, so they are filed under the instruct member. Apache 2.0 throughout. The Hub shows XHToken (fullname SparkLLM) as an unverified organisation. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Spark-X2.6-4B Spark-X2.6-4B is SparkLLM's open-weight 4B text model, Spark-X2.5-4B's architecture with new post-training, a 1M-token context and three thinking levels, on Hugging Face with FP8, INT8, INT4 and GGUF builds under Apache 2.0. | instruct | 4B (Hub safetensors count 4.11B) | 2026-10-19 | apache-2.0 | Our listing (C) · Download · Spark-X2.6-4B-GGUF · Spark-X2.6-4B-FP8 · Spark-X2.6-4B-INT8 · Spark-X2.6-4B-INT4 | |
| Spark-X2.6-4B-Base | base | 4B (Hub safetensors count 4.11B) | 2026-10-19 | apache-2.0 | Download | |
| Spark-X2.5-4B-Base | base | 4B (Hub safetensors count 4.11B) | 2026-08-28 | apache-2.0 | Download | |
| Spark-X2.5-1.7B-Base | base | 1.7B (Hub safetensors count 1.71B) | 2026-08-28 | apache-2.0 | Download | |
| Spark-X2.5-4B | instruct | 4B (Hub safetensors count 4.11B) | 2026-08-24 | apache-2.0 | Download · Spark-X2.5-4B-GGUF · Spark-X2.5-4B-INT8 · Spark-X2.5-4B-FP8 | |
| Spark-X2.5-1.7B | instruct | 1.7B (Hub safetensors count 1.71B) | 2026-08-24 | apache-2.0 | Download · Spark-X2.5-1.7B-GGUF · Spark-X2.5-1.7B-INT8 · Spark-X2.5-1.7B-FP8 |
Swiss AI Initiative (EPFL, ETH Zurich, CSCS)
Publishes from huggingface.co/swiss-ai. Checked 2026-10-21.
Apertus (v1 2509, v1.1, v1.5) from 2025-08-13
One major version so far. Apertus 8B and 70B (2509, September 2025, base and instruct); v1.1 small models (0.5B, 1.5B, 4B, February to May 2026, base and instruct, with Swiss AI's own MLX and NVFP4A16 builds); and v1.5 (24 July 2026), which adds vision (image-text-to-text), with base, SFT, DPO, RL and final checkpoints. Every v1.5 repository is gated with automatic approval behind a form that asks for company and institutional email. Apache 2.0 throughout. A toxicity classifier and a WavTokenizer checkpoint in the organisation are left out. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Apertus-v1.5-8B Apertus v1.5 8B is the Swiss AI Initiative's open-data, open-weight multilingual model of 8.9B parameters that takes text, images and audio and answers in text, with optional thinking and a 262,144-token context. The download is gated behind a click-through. | instruct (vision-language) | 8B (Hub safetensors count 8.90B) | 2026-07-24 | apache-2.0 | Our listing (C) · Download (gated) | |
| Apertus-v1.5-8B-base | base | 8B (Hub safetensors count 8.84B) | 2026-09-23 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-70B-base | base | 70B (Hub safetensors count 71.94B) | 2026-09-23 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-8B-SFT | instruct (SFT stage) | 8B (Hub safetensors count 8.90B) | 2026-08-22 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-8B-RL | instruct (RL stage) | 8B (Hub safetensors count 8.90B) | 2026-08-22 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-8B-DPO | instruct (DPO stage) | 8B (Hub safetensors count 8.90B) | 2026-08-22 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-70B-SFT | instruct (SFT stage) | 70B (Hub safetensors count 72.01B) | 2026-08-22 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-70B-RL | instruct (RL stage) | 70B (Hub safetensors count 72.01B) | 2026-08-22 | apache-2.0 | Download (gated) | |
| Apertus-v1.5-70B | instruct (vision-language) | 70B (Hub safetensors count 72.01B) | 2026-07-24 | apache-2.0 | Download (gated) | |
| Apertus-v1.1-1.5B-Instruct | instruct | 1.5B (Hub safetensors count 1.51B) | 2026-05-03 | apache-2.0 | Download · Apertus-v1.1-1.5B-Instruct-MLX-INT3 · Apertus-v1.1-1.5B-Instruct-MLX-INT4 · Apertus-v1.1-1.5B-Instruct-MLX-INT6 · Apertus-v1.1-1.5B-Instruct-vLLM-NVFP4A16 | |
| Apertus-v1.1-4B-Instruct | instruct | 4B (Hub safetensors count 3.83B) | 2026-05-01 | apache-2.0 | Download · Apertus-v1.1-4B-Instruct-MLX-INT3 · Apertus-v1.1-4B-Instruct-MLX-INT4 · Apertus-v1.1-4B-Instruct-MLX-INT6 | |
| Apertus-v1.1-0.5B-Instruct | instruct | 0.5B (Hub safetensors count 573M) | 2026-05-01 | apache-2.0 | Download · Apertus-v1.1-0.5B-Instruct-MLX-INT3 · Apertus-v1.1-0.5B-Instruct-MLX-INT4 · Apertus-v1.1-0.5B-Instruct-MLX-INT6 | |
| Apertus-v1.1-4B | base | 4B (Hub safetensors count 3.83B) | 2026-04-12 | apache-2.0 | Download | |
| Apertus-v1.1-0.5B | base | 0.5B (Hub safetensors count 438M) | 2026-03-30 | apache-2.0 | Download | |
| Apertus-v1.1-1.5B | base | 1.5B (Hub safetensors count 1.51B) | 2026-02-18 | apache-2.0 | Download | |
| Apertus-8B-2509 | base | 8B (Hub safetensors count 8.05B) | 2025-09-02 | apache-2.0 | Download | |
| Apertus-70B-2509 | base | 70B (Hub safetensors count 70.60B) | 2025-09-02 | apache-2.0 | Download | |
| Apertus-70B-Instruct-2509 | instruct | 70B (Hub safetensors count 70.60B) | 2025-09-01 | apache-2.0 | Download | |
| Apertus-8B-Instruct-2509 | instruct | 8B (Hub safetensors count 8.05B) | 2025-08-13 | apache-2.0 | Download |
Tencent (Hunyuan / Hy)
Publishes from huggingface.co/tencent. Checked 2026-10-21.
Hy3 from 2026-04-23
Hy3 preview (open-sourced 2026-04-23 per its card) and Hy3 share the 295B/21B shape; Hy3's date is our listing's firstReleased (repository created 2026-07-02). The previews are under the Tencent Hy Community License, Hy3 under Apache-2.0. Hy3's card also points to a GGUF at AngelSlim/Hy3-GGUF, Tencent's AngelSlim toolkit's Hugging Face organisation; it is not in the tencent organisation, so it isn't listed as a build.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hy3-preview-Base | base | 295B total, 21B activated (+3.8B MTP) | 2026-04-23 | other (Tencent Hy Community License; excludes EU, UK, South Korea) | Download | |
| Hy3-preview | instruct | 295B total, 21B activated (+3.8B MTP) | 2026-04-23 | other (Tencent Hy Community License; excludes EU, UK, South Korea) | Download | |
| Hy3 Hy3 is Tencent's open-weight mixture-of-experts language model, 295B parameters with 21B active, a 256K-token context and per-request reasoning effort, published on Hugging Face in BF16 and FP8 under Apache 2.0. | instruct | 295B total, 21B activated (+3.8B MTP) | 2026-07-06 | apache-2.0 | Our listing (B) · Download · Hy3-FP8 |
Hy4 from 2026-08-27
Newer than our listed family and still a preview; no listing. Date is the Hub repository creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hy4-preview | instruct | 770B total, 49B activated | 2026-08-27 | apache-2.0 | Download · Hy4-preview-FP8 |
Older versions (7 families, 22 models)
Hy-Memory from 2026-10-15
Specialised side line, newer than Hy3: a memory-writer model for external memory frameworks. Hub repository creation date.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hy-Memory-30B-A3B | instruct | ~30B total, ~3B active | 2026-10-15 | apache-2.0 | Download · Hy-Memory-30B-A3B-FP8 · Hy-Memory-30B-A3B-GGUF |
Hy-MT2 from 2026-05-21
Translation line, open-sourced 2026-05-21 per the card.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hy-MT2-1.8B | instruct | 1.8B | 2026-05-21 | apache-2.0 | Download · Hy-MT2-1.8B-FP8 · Hy-MT2-1.8B-GGUF · Hy-MT2-1.8B-2Bit-GGUF · Hy-MT2-1.8B-1.25Bit-GGUF | |
| Hy-MT2-7B | instruct | 7B | 2026-05-21 | apache-2.0 | Download · Hy-MT2-7B-FP8 · Hy-MT2-7B-GGUF | |
| Hy-MT2-30B-A3B | instruct | 30B total, 3B active | 2026-05-21 | apache-2.0 | Download · Hy-MT2-30B-A3B-FP8 · Hy-MT2-30B-A3B-GGUF |
Hunyuan-MT (MT1 and HY-MT1.5) from 2025-09-01
Translation line. Hunyuan-MT open-sourced 2025-09-01 and HY-MT1.5 2025-12-30, per the cards. The 2bit and 1.25bit HY-MT1.5-1.8B repositories (2026-04-29) are listed as builds.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hunyuan-MT-7B | instruct | 7B | 2025-09-01 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download · Hunyuan-MT-7B-fp8 | |
| Hunyuan-MT-Chimera-7B | instruct (ensemble) | 7B | 2025-09-01 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download · Hunyuan-MT-Chimera-7B-fp8 | |
| HY-MT1.5-1.8B | instruct | 1.8B | 2025-12-30 | other (Tencent HY Community License; excludes EU, UK, South Korea) | Download · HY-MT1.5-1.8B-FP8 · HY-MT1.5-1.8B-GPTQ-Int4 · HY-MT1.5-1.8B-GGUF · Hy-MT1.5-1.8B-2bit · Hy-MT1.5-1.8B-2bit-GGUF · Hy-MT1.5-1.8B-1.25bit · Hy-MT1.5-1.8B-1.25bit-GGUF | |
| HY-MT1.5-7B | instruct | 7B | 2025-12-30 | other (Tencent HY Community License; excludes EU, UK, South Korea) | Download · HY-MT1.5-7B-FP8 · HY-MT1.5-7B-GPTQ-Int4 · HY-MT1.5-7B-GGUF |
Hunyuan dense (0.5B to 7B) from 2025-07-30
Open-sourced 2025-07-30 per the cards (licence files say release date 4 August 2025). Hub metadata carries no licence tag; the LICENSE files are the Tencent Hunyuan Community License.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hunyuan-0.5B-Pretrain | base | 0.5B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download | |
| Hunyuan-0.5B-Instruct | instruct | 0.5B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | tencent/Hunyuan-0.5B-Pretrain | Download · Hunyuan-0.5B-Instruct-FP8 · Hunyuan-0.5B-Instruct-GPTQ-Int4 · Hunyuan-0.5B-Instruct-AWQ-Int4 |
| Hunyuan-1.8B-Pretrain | base | 1.8B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download | |
| Hunyuan-1.8B-Instruct | instruct | 1.8B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download · Hunyuan-1.8B-Instruct-FP8 · Hunyuan-1.8B-Instruct-GPTQ-Int4 · Hunyuan-1.8B-Instruct-AWQ-Int4 | |
| Hunyuan-4B-Pretrain | base | 4B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download | |
| Hunyuan-4B-Instruct | instruct | 4B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | tencent/Hunyuan-4B-Pretrain | Download · Hunyuan-4B-Instruct-FP8 · Hunyuan-4B-Instruct-GPTQ-Int4 · Hunyuan-4B-Instruct-AWQ-Int4 |
| Hunyuan-7B-Pretrain | base | 7B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | Download | |
| Hunyuan-7B-Instruct | instruct | 7B | 2025-07-30 | other (Tencent Hunyuan Community License; excludes EU, UK, South Korea) | tencent/Hunyuan-7B-Pretrain | Download · Hunyuan-7B-Instruct-FP8 · Hunyuan-7B-Instruct-GPTQ-Int4 · Hunyuan-7B-Instruct-AWQ-Int4 |
Hunyuan-A13B from 2025-06-27
Open-sourced 2025-06-27 per the card and licence.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hunyuan-A13B-Pretrain | base | 80B total, 13B active | 2025-06-27 | other (Tencent Hunyuan A13B Community License) | Download | |
| Hunyuan-A13B-Instruct | instruct | 80B total, 13B active | 2025-06-27 | other (Tencent Hunyuan A13B Community License) | Download · Hunyuan-A13B-Instruct-FP8 · Hunyuan-A13B-Instruct-GPTQ-Int4 · Hunyuan-A13B-Instruct-GGUF |
Hunyuan-7B (0124) from 2025-01-24
Open-sourced 2025-01-24 per the card. Superseded by the July 2025 dense line but still hosted.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hunyuan-7B-Pretrain-0124 | base | 7B | 2025-01-24 | other (tencent-license) | Download | |
| Hunyuan-7B-Instruct-0124 | instruct | 7B | 2025-01-24 | other (tencent-license) | Download |
Hunyuan-Large from 2024-11-05
Dated by the licence's release date (repository created 2024-10-22). One repository holds three checkpoints in folders; licence excludes the EU.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Hunyuan-A52B-Pretrain | base | 389B total, 52B active | 2024-11-05 | other (Tencent Hunyuan Community License; excludes EU) | Download | |
| Hunyuan-A52B-Instruct | instruct | 389B total, 52B active | 2024-11-05 | other (Tencent Hunyuan Community License; excludes EU) | Download · Hunyuan-A52B-Instruct-FP8 |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Hy3 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
The Endless Frontier (endless-frontier)
Publishes from huggingface.co/endless-frontier. Checked 2026-10-21.
BigBang from 2026-08-02
The Endless Frontier's general research model. Only a BF16 repository from the publisher; the GGUFs that carry its downloads are others' (bartowski's above all). The organisation isn't verified and names no legal entity or website.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| BigBang-v1 BigBang-v1 is The Endless Frontier's general-purpose fine-tune of Qwen3.6-35B-A3B (35B mixture of experts, 3B active), post-trained on about 10,000 synthetic research and technical problems, published on Hugging Face in BF16 under Apache 2.0. | reasoning (scientific research, search, coding) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-08-02 | apache-2.0 | Qwen/Qwen3.6-35B-A3B | Our listing (C) · Download |
Older versions (1 family, 2 models)
Fx from 2026-09-19
A separate, later line of domain models, not an earlier version of BigBang.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Fx-Bio (The card calls it Fx-Bio-0913. DeepseekV4ForCausalLM architecture. The card names the base as 'DeepSeek-V4-Flash' without an organisation; the repository id is our reading.) | reasoning (biomedical) | 284B (DeepSeek-V4-Flash size; MoE; 284,334,567,511 by the Hub's safetensors count) | 2026-09-19 | apache-2.0 | deepseek-ai/DeepSeek-V4-Flash (per the card's text; not declared in the metadata) | Download |
| Fx-Work | reasoning (professional file-and-tool work, multimodal) | 35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count) | 2026-09-29 | apache-2.0 | Qwen/Qwen3.6-35B-A3B | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: BigBang-v1 3. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Thinking Machines Lab
Publishes from huggingface.co/thinkingmachines, github.com/thinking-machines-lab. Checked 2026-10-21.
Inkling from 2026-07-14
Thinking Machines Lab's only open-weight family: Inkling and Inkling-Small, both multimodal (text, image and audio in, text out) and Apache 2.0, each with the lab's own NVFP4 build. No base checkpoints were released. The organisation also holds two Llama 3 tokenizer repositories (no weights), left out. The Hub shows thinkingmachines as an unverified organisation; Tinker (data/tools/tinker.json) serves both models, which supports it, but thinkingmachines.ai was not re-read here. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Inkling-Small Inkling-Small is Thinking Machines Lab's open-weight mixture-of-experts model, 276B total and 12B active parameters, taking text, images and audio and returning text, with adjustable thinking effort, published in BF16 and NVFP4 under Apache 2.0. | instruct (multimodal: text, image and audio in, text out; adjustable thinking effort) | 265.96B (Hub safetensors count; the name states none) | 2026-07-27 | apache-2.0 | Our listing (B) · Download · Inkling-Small-NVFP4 | |
| Inkling Inkling is Thinking Machines Lab's large open-weight mixture-of-experts model, 975B total and 41B active parameters, taking text, images and audio and returning text, with adjustable thinking effort, published in BF16 and NVFP4 under Apache 2.0. | instruct (multimodal: text, image and audio in, text out; adjustable thinking effort) | 952.38B (Hub safetensors count; the name states none) | 2026-07-14 | apache-2.0 | Our listing (B) · Download · Inkling-NVFP4 |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Inkling 1, Inkling-Small 3. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
UkisAI (ukisai)
Publishes from huggingface.co/ukisai. Checked 2026-10-21.
Swift 1 from 2026-09-08
UkisAI's Swift line, reasoning-efficient post-trains of other labs' models. Swift 1.0 (published as Swift-Qwen3.8-27b) and Swift 1.5 are in one major family by the decimal-is-minor rule. ukisai is a Hugging Face user account, not an organisation; ukisai.com's Swift page links to it. Swift-licensed repositories use UkisAI's own Swift Open License 1.0 (commercial use only below US$1M gross annual revenue) on UkisAI's changes; the base's licence still applies to the base. Member dates are the Hub's createdAt unless our listing records the public release.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| Swift-Qwen3.8-27b (Swift 1.0, still published; the batch skips it behind Swift 1.5.) | reasoning (fewer thinking tokens; Swift 1.0) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-08 | other (swift-open-license-1.0) | Qwen/Qwen3.8-27B | Download · Swift-Qwen3.8-27B-GGUF · Swift-Qwen3.8-27B-NVFP4 · Swift-Qwen3.8-27b-int4-AMD · Swift-Qwen3.8-27b-BF16-AMD |
| Swift-1.5-Qwen3.8-27b Swift 1.5 Qwen3.8-27B is UkisAI's post-trained Qwen3.8-27B that aims to think in fewer tokens, a 27B dense vision-language model with the base's 262,144-token context, published in BF16 with GGUF, AWQ, NVFP4, FP8 and MLX builds under UkisAI's own licence. | reasoning (fewer thinking tokens) | 27B (27,781,427,952 by the Hub's safetensors count) | 2026-09-21 | other (swift-open-license-1.0) | Qwen/Qwen3.8-27B | Our listing (C) · Download · Swift-1.5-Qwen3.8-27B-GGUF · Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF · Swift-1.5-Qwen3.8-27b-W4A16-AWQ · Swift-1.5-Qwen3.8-27b-W4A16-AutoRound · Swift-1.5-Qwen3.8-27b-INT4 · Swift-1.5-Qwen3.8-27b-NVFP4 · Swift-1.5-Qwen3.8-27b-Quark-FP8-dynamic-AMD · Swift-1.5-4bit-MLX · Swift-1.5-5bit-MLX · Swift-1.5-3bit-MLX-TextOnly · Swift-1.5-Qwen3.8-27b-ONNX-BF16 · Swift-1.5-Qwen3.8-27b-ONNX-INT8 · Swift-1.5-Qwen3.8-27b-ONNX-INT4 · Swift-1.5-Qwen3.8-27b-ONNX-INT4-Mixed · Swift-1.5-Qwen3.8-27b-ONNX-NVFP4 |
| Swift1.5-Qwen3.8-Flash-Next Swift 1.5 Qwen3.8-Flash-Next is UkisAI's post-trained Qwen3.8-Flash-Next that aims to think in fewer tokens, a mixture-of-experts vision-language model of 180B parameters in its files with a 262,144-token context, published in BF16 with GGUF, INT4 and NVFP4 builds. (Renamed from ukisai/Swift-Qwen3.8-Flash-Next, which still redirects; most of its builds' metadata still name the old repository. Base licence is the Qwen Community License 1.0. Released 2026-09-21 per our listing; the Hub's createdAt is 2026-09-22.) | reasoning (fewer thinking tokens) | 125B (Qwen3.8-Flash-Next's stated size; MoE; 179,999,981,459 by the Hub's safetensors count) | 2026-09-21 | other (swift-open-license-1.0) | Qwen/Qwen3.8-Flash-Next | Our listing (D) · Download · Swift-1.5-Qwen3.8-Flash-Next-GGUF · Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF · Swift-1.5-Qwen3.8-Flash-Next-W4A16-AWQ · Swift-1.5-Qwen3.8-Flash-Next-W4A16-AutoRound · Swift-1.5-Qwen3.8-Flash-Next-NVFP4 |
| Swift-1.5-Ornith-9B (Gated (manual approval): the files open after accepting the card's conditions. Built on Ornith-1.5-9B, itself built on Qwen3.5 (see ornith.json). Apache-2.0, unlike the Qwen3.8 Swifts.) | reasoning (fewer thinking tokens) | 9B (9,653,104,368 by the Hub's safetensors count) | 2026-10-09 | apache-2.0 | ornith-ai/Ornith-1.5-9B | Download · Swift-1.5-Ornith-9B-GGUF |
| Swift-Bonsai-2-GGUF (Published only as GGUF; no BF16 repository.) | reasoning (fewer thinking tokens; 2-bit GGUF only) | 27B (26,895,998,464 by the Hub's GGUF count) | 2026-09-22 | apache-2.0 | prism-ml/Ternary-Bonsai-2-27B-gguf (prism-ml's ternary build of Qwen/Qwen3.8-27B) | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: Swift 1.5 Qwen3.8-27B 5, Swift 1.5 Qwen3.8-Flash-Next 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
VieNeu (pnnbao-ump)
Publishes from huggingface.co/pnnbao-ump, github.com/pnnbao97/VieNeu-TTS. Checked 2026-10-21.
VieNeu-TTS v3 from 2026-06-05
Vietnamese and English text-to-speech from an individual's account (Phạm Nguyễn Ngọc Bảo, per the batch research notes). v3 Turbo is trained from scratch on a Qwen3-shaped architecture with the publisher's own small config (no base declared). v3 Nano (4 September 2026) shows no safetensors count on the Hub. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| VieNeu-TTS-v3-Turbo VieNeu-TTS v3 Turbo is an individual developer's open-weight Vietnamese and English text-to-speech model at 48 kHz, about 131M parameters trained from scratch on a Qwen3-style backbone, with three-region preset voices, voice cloning and an OpenAI-compatible server in its vieneu SDK. | text-to-speech | 131M (Hub safetensors count; the name states none) | 2026-06-05 | apache-2.0 | Our listing (C) · Download | |
| VieNeu-TTS-v3-Nano | text-to-speech | not stated (no safetensors count on the Hub) | 2026-09-04 | apache-2.0 | Download |
Older versions (4 families, 7 models)
VieNeu-TTS v2 from 2026-03-31
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| VieNeu-TTS-v2 | text-to-speech | 294M (Hub safetensors count; the name states none) | 2026-05-06 | apache-2.0 | Download | |
| VieNeu-TTS-v2-Turbo | text-to-speech | 112M (Hub safetensors count; the name states none) | 2026-03-31 | apache-2.0 | Download · VieNeu-TTS-v2-Turbo-GGUF |
MedCrab from 2025-11-28
A medical translation fine-tune of Qwen/Qwen2.5-1.5B by the same account, not speech.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MedCrab-1.5B | translation (medical) | 1.5B (Hub safetensors count 1.54B) | 2025-11-28 | cc-by-nc-4.0 | Qwen/Qwen2.5-1.5B | Download |
kani-tts-370m-vie from 2025-11-10
A Vietnamese fine-tune of nineninesix/kani-tts-370m by the same account.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| kani-tts-370m-vie | text-to-speech | 370M (Hub safetensors count 370M) | 2025-11-10 | apache-2.0 | nineninesix/kani-tts-370m | Download |
VieNeu-TTS (v1: 0.5B on NeuTTS Air, and 0.3B) from 2025-10-28
The first VieNeu-TTS is a fine-tune of neuphonic/neutts-air; VieNeu-TTS-0.3B (January 2026) declares no base. Its q4 and q8 GGUF builds are cc-by-nc-4.0 while the safetensors repository is apache-2.0. A LoRA adapter (VieNeu-TTS-0.3B-lora-ngoc-huyen) is left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| VieNeu-TTS-0.3B-ngoc-huyen | text-to-speech (single-voice fine-tune, ngoc-huyen) | 0.3B (Hub safetensors count 244M) | 2026-05-02 | apache-2.0 | Download · VieNeu-TTS-0.3B-ngoc-huyen-gguf-Q4_0 | |
| VieNeu-TTS-0.3B | text-to-speech | 0.3B (Hub safetensors count 244M) | 2026-01-07 | apache-2.0 | Download · VieNeu-TTS-0.3B-q4-gguf · VieNeu-TTS-0.3B-q8-gguf | |
| VieNeu-TTS | text-to-speech | 553M (Hub safetensors count; the name states none) | 2025-10-28 | apache-2.0 | neuphonic/neutts-air | Download · VieNeu-TTS-q4-gguf · VieNeu-TTS-q8-gguf |
Xiaomi (MiMo)
Publishes from huggingface.co/XiaomiMiMo. Checked 2026-10-21.
MiMo-V2 (V2, V2.5, V2.6) from 2025-12-16
Xiaomi's second-generation MiMo: V2-Flash (December 2025, 309.8B by the Hub count), V2.5 and V2.5-Pro (April 2026, 310.8B and 1,023B), V2.5-ASR, and V2.6 Flash and Pro (September 2026) in RL and MOPD editions, plus MiMo-V2.6-Distill-Qwen-9B, a supervised fine-tune of Qwen/Qwen3.5-9B on MiMo-V2.6 outputs (declared in its card). V2.5 and V2.6 are treated as minor versions of V2. The two DFlash repositories hold the full-size weights (FP4 for Pro) with a drafter and are listed as builds. MiMo-V2-Flash-Base's card tags itself as quantised from itself. Xiaomi publishes no GGUF for the V2 line. MIT throughout. Dates are the repository createdAt from the Hub API (repositories are often created privately before release) unless the member says otherwise.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiMo-V2.6-Distill-Qwen-9B MiMo-V2.6 Distill Qwen 9B is Xiaomi's open-weight 9B agentic model, Qwen3.5-9B fine-tuned on data generated by MiMo-V2.6 for coding, agent, visual coding and cybersecurity tasks, released in BF16 under MIT as a starting point for RL research. | instruct (vision-language; supervised fine-tune of Qwen3.5-9B on MiMo-V2.6 outputs) | 9B (Hub safetensors count 9.41B) | 2026-09-21 | mit | Qwen/Qwen3.5-9B | Our listing (C) · Download |
| MiMo-V2.6-Pro-MOPD | post-trained (MOPD stage, per name; card not read) | 1024.22B (Hub safetensors count; the name states none) | 2026-09-27 | mit | Download | |
| MiMo-V2.6-Flash-MOPD | post-trained (MOPD stage, per name; card not read) | 310.76B (Hub safetensors count; the name states none) | 2026-09-27 | mit | Download | |
| MiMo-V2.6-Pro-RL | reasoning (RL stage, per name) | 1024.22B (Hub safetensors count; the name states none) | 2026-09-21 | mit | Download | |
| MiMo-V2.6-Flash-RL | reasoning (RL stage, per name) | 310.76B (Hub safetensors count; the name states none) | 2026-09-21 | mit | Download | |
| MiMo-V2.5-Pro-Base | base | 1023.18B (Hub safetensors count; the name states none) | 2026-04-27 | mit | Download | |
| MiMo-V2.5-Pro | instruct | 1023.18B (Hub safetensors count; the name states none) | 2026-04-27 | mit | Download · MiMo-V2.5-Pro-FP4-DFlash | |
| MiMo-V2.5-Base | base | 310.76B (Hub safetensors count; the name states none) | 2026-04-27 | mit | Download | |
| MiMo-V2.5 | instruct | 310.76B (Hub safetensors count; the name states none) | 2026-04-27 | mit | Download · MiMo-V2.5-DFlash | |
| MiMo-V2.5-ASR | speech-to-text | 7.62B (Hub safetensors count; the name states none) | 2026-04-23 | mit | Download | |
| MiMo-V2-Flash-Base | base | 309.77B (Hub safetensors count; the name states none) | 2025-12-16 | mit | Download | |
| MiMo-V2-Flash | instruct | 309.77B (Hub safetensors count; the name states none) | 2025-12-16 | mit | Download |
Older versions (4 families, 12 models)
MiMo-Embodied from 2025-11-19
A vision-language model for embodied and driving tasks.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiMo-Embodied-7B | instruct (vision-language) | 7B (Hub safetensors count 8.31B) | 2025-11-19 | mit | Download |
MiMo-Audio from 2025-09-18
Audio language model (audio in and out), base and instruct. Its audio tokenizer repository is left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiMo-Audio-7B-Instruct | speech and audio (audio in, text or audio out) | 7B (Hub safetensors count 8.02B) | 2025-09-18 | mit | Download | |
| MiMo-Audio-7B-Base | base | 7B (Hub safetensors count 8.02B) | 2025-09-18 | mit | Download |
MiMo-VL from 2025-05-30
Vision-language models on the MiMo-7B line, SFT and RL editions, with a 2508 update.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiMo-VL-7B-SFT-2508 | instruct (vision-language, SFT, 2508 update) | 7B (Hub safetensors count 8.31B) | 2025-08-07 | mit | Download | |
| MiMo-VL-7B-RL-2508 | reasoning (vision-language, RL, 2508 update) | 7B (Hub safetensors count 8.31B) | 2025-08-07 | mit | Download | |
| MiMo-VL-7B-SFT | instruct (vision-language, SFT) | 7B (Hub safetensors count 8.31B) | 2025-05-30 | mit | Download · MiMo-VL-7B-SFT-GGUF | |
| MiMo-VL-7B-RL | reasoning (vision-language, RL) | 7B (Hub safetensors count 8.31B) | 2025-05-30 | mit | Download · MiMo-VL-7B-RL-GGUF |
MiMo (7B) from 2025-04-29
Xiaomi's first open reasoning models (April 2025): base, SFT, RL-Zero and RL, plus a May 2025 RL update. A separate MTP-layers repository is left out.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| MiMo-7B-RL-0530 | reasoning (RL, 0530 update) | 7B (Hub safetensors count 7.83B) | 2025-05-30 | mit | Download | |
| MiMo-7B-SFT | instruct (SFT stage) | 7B (Hub safetensors count 7.83B) | 2025-04-29 | mit | Download | |
| MiMo-7B-RL-Zero | reasoning (RL-Zero, RL straight from base) | 7B (Hub safetensors count 7.83B) | 2025-04-29 | mit | Download | |
| MiMo-7B-RL | reasoning (RL) | 7B (Hub safetensors count 7.83B) | 2025-04-29 | mit | Download | |
| MiMo-7B-Base | base | 7B (Hub safetensors count 7.83B) | 2025-04-29 | mit | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: MiMo-V2.6 Distill Qwen 9B 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
Z.ai (Zhipu AI)
Publishes from huggingface.co/zai-org, huggingface.co/THUDM. Checked 2026-10-21.
GLM-5 from 2026-02-11
Z.ai's GitHub table (zai-org/GLM-5) lists GLM-5, 5.1, 5.2, 5.3 and 5.3-Flash with ModelScope mirrors under ZhipuAI/. No base editions published. Dates are Hub repository creation dates, except GLM-5.3 (first commit 2026-08-27; repo created 2026-08-25) and GLM-5.3-Flash (first public card edits 2026-08-26), which follow our listings.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GLM-5 (Hybrid reasoning; text only) | instruct | 744B-A40B | 2026-02-11 | mit | Download · GLM-5-FP8 | |
| GLM-5.1 | instruct | 744B-A40B | 2026-04-03 | mit | Download · GLM-5.1-FP8 | |
| GLM-5.2 | instruct | 744B-A40B | 2026-06-16 | mit | Download · GLM-5.2-FP8 | |
| GLM-5.3 GLM-5.3 is Z.ai's open-weight mixture-of-experts model for coding and long-horizon agent work, 744B parameters with 40B active and a 1M-token context, published on Hugging Face in FP8 and BF16 under its own licence. (The main repository holds the FP8 weights; BF16 is the separate build. Thinking can't be turned off (reasoning_effort low/high/max)) | reasoning | 744B-A40B | 2026-08-27 | other (GLM-5.3 License) | same base model as zai-org/GLM-5.2 (card: every gain comes from post-training) | Our listing (B) · Download · GLM-5.3-BF16 |
| GLM-5.3-Flash GLM-5.3-Flash is Z.ai's open-weight multimodal mixture-of-experts model, 320B parameters with 18B active and a 1,048,576-token context, published on Hugging Face in FP8 and BF16 under the MIT licence. (Natively multimodal (image-text-to-text); card says it starts from a newly trained base model, not named. Main repository is FP8) | instruct | 320B-A18B | 2026-08-26 | mit | Our listing (B) · Download · GLM-5.3-Flash-BF16 |
GLM-4 from 2024-06-04
Everything Z.ai numbers 4.x: GLM-4-9B (2024), the 0414 series with its GLM-Z1 reasoning editions, GLM-4.1V, GLM-4.5/4.5V, GLM-4.6/4.6V and GLM-4.7. GLM-4.5 onward is a different MoE architecture from GLM-4-9B, so it could be read as its own major version; see open. Research fine-tunes on GLM-4-9B (LongWriter-glm4-9b, LongCite-glm4-9b, LongReward-glm4-9b-DPO, webrl-glm-4-9b, androidgen-glm-4-9b, codegeex4-all-9b) and GLM-4.1V-9B-Base (AutoGLM-Phone-9B, Glyph, UI2Code_N, WebVIA-Agent) are left out as separate products. glm-4-voice-9b (speech) is left out as out of scope. Dates are Hub repository creation dates, apart from GLM-4.7-Flash, which follows our listing.
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GLM-4-9B (glm-4-9b-hf is the publisher's Transformers-format copy (2025-01-16)) | base | 9B | 2024-06-04 | other (glm-4 licence) | Download · glm-4-9b-hf | |
| GLM-4-9B-Chat | instruct | 9B | 2024-06-04 | other (glm-4 licence) | Download · glm-4-9b-chat-hf | |
| GLM-4-9B-Chat-1M (1M-token context edition) | instruct | 9B | 2024-06-04 | other (glm-4 licence) | Download · glm-4-9b-chat-1m-hf | |
| GLM-4V-9B (Vision-language; Hub counts 13.9B parameters including the vision encoder) | instruct | 9B | 2024-06-04 | other (glm-4 licence) | Download | |
| GLM-4-9B-0414 | instruct | 9B | 2025-04-07 | mit | Download | |
| GLM-4-32B-Base-0414 | base | 32B | 2025-04-07 | mit | Download | |
| GLM-4-32B-0414 | instruct | 32B | 2025-04-07 | mit | Download | |
| GLM-Z1-9B-0414 (Card titles it GLM-4-Z1-9B-0414) | reasoning | 9B | 2025-04-08 | mit | Download | |
| GLM-Z1-32B-0414 | reasoning | 32B | 2025-04-08 | mit | zai-org/GLM-4-32B-0414 | Download |
| GLM-Z1-Rumination-32B-0414 (Deep-research style reasoning edition; card titles it GLM-4-Z1-Rumination-32B-0414) | reasoning | 32B | 2025-04-13 | mit | Download | |
| GLM-4.1V-9B-Base (Vision-language) | base | 9B | 2025-06-28 | mit | zai-org/GLM-4-9B-0414 | Download |
| GLM-4.1V-9B-Thinking (Vision-language) | reasoning | 9B | 2025-06-28 | mit | zai-org/GLM-4-9B-0414 | Download |
| GLM-4.5-Base | base | 355B-A32B | 2025-07-20 | mit | Download | |
| GLM-4.5 (Hybrid reasoning) | instruct | 355B-A32B | 2025-07-20 | mit | Download · GLM-4.5-FP8 | |
| GLM-4.5-Air-Base | base | 106B-A12B | 2025-07-20 | mit | Download | |
| GLM-4.5-Air GLM-4.5-Air is Z.ai's open-weight mixture-of-experts model for English and Chinese, 106B total and 12B active parameters, with thinking that turns off per request, tool calling and a 128K context, published in BF16 and FP8 under MIT. (Hybrid reasoning) | instruct | 106B-A12B | 2025-07-20 | mit | Our listing (C) · Download · GLM-4.5-Air-FP8 | |
| GLM-4.5V (Vision-language) | instruct | 106B-A12B | 2025-08-10 | mit | zai-org/GLM-4.5-Air-Base | Download · GLM-4.5V-FP8 |
| GLM-4.6 | instruct | 355B-A32B | 2025-09-29 | mit | Download · GLM-4.6-FP8 | |
| GLM-4.6V (Vision-language) | instruct | 106B | 2025-12-07 | mit | Download · GLM-4.6V-FP8 | |
| GLM-4.6V-Flash (Vision-language) | instruct | 9B | 2025-12-07 | mit | Download | |
| GLM-4.7 | instruct | 355B-A32B | 2025-12-22 | mit | Download · GLM-4.7-FP8 | |
| GLM-4.7-Flash GLM-4.7-Flash is Z.ai's open-weight mixture-of-experts reasoning model, 30B parameters with about 3B active per token, for English and Chinese, published on Hugging Face in BF16 under the MIT licence. (Thinking on by default; no publisher quantised build) | instruct | 30B-A3B | 2026-01-19 | mit | Our listing (B) · Download |
Older versions (7 families, 19 models)
GLM-Edge from 2024-11-20
Small on-device line under the glm-4 licence; Z.ai names it separately from GLM-4
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| glm-edge-1.5b-chat | instruct | 1.5B | 2024-11-20 | other (glm-4 licence) | Download · glm-edge-1.5b-chat-gguf | |
| glm-edge-4b-chat | instruct | 4B | 2024-11-20 | other (glm-4 licence) | Download · glm-edge-4b-chat-gguf | |
| glm-edge-v-2b (Vision-language) | instruct | 2B | 2024-11-24 | other (glm-4 licence) | Download · glm-edge-v-2b-gguf | |
| glm-edge-v-5b (Vision-language) | instruct | 5B | 2024-11-24 | other (glm-4 licence) | Download · glm-edge-v-5b-gguf |
ChatGLM3 from 2023-10-25
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ChatGLM3-6B | instruct | 6B | 2023-10-25 | other (ChatGLM model licence; code apache-2.0) | Download | |
| ChatGLM3-6B-Base | base | 6B | 2023-10-26 | other (ChatGLM model licence; code apache-2.0) | Download | |
| ChatGLM3-6B-32K | instruct | 6B | 2023-10-26 | other (ChatGLM model licence; code apache-2.0) | Download | |
| ChatGLM3-6B-128K | instruct | 6B | 2024-01-30 | other (ChatGLM model licence; code apache-2.0) | THUDM/chatglm3-6b | Download |
ChatGLM2 from 2023-06-24
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ChatGLM2-6B (Card: commercial use allowed after a registration questionnaire) | instruct | 6B | 2023-06-24 | other (ChatGLM model licence; code apache-2.0) | Download · chatglm2-6b-int4 | |
| ChatGLM2-6B-32K | instruct | 6B | 2023-07-30 | other (ChatGLM model licence; code apache-2.0) | THUDM/chatglm2-6b | Download · chatglm2-6b-32k-int4 |
VisualGLM from 2023-05-17
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| VisualGLM-6B (Vision-language) | instruct | 6B | 2023-05-17 | other (ChatGLM model licence; code apache-2.0) | Download |
ChatGLM from 2023-03-13
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| ChatGLM-6B (Card: commercial use allowed after a registration questionnaire; weights in .bin format) | instruct | 6.2B | 2023-03-13 | other (ChatGLM model licence; code apache-2.0) | Download · chatglm-6b-int4 · chatglm-6b-int8 · chatglm-6b-int4-qe |
WebGLM from 2023-05-28
Web-search question-answering models
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| WebGLM | instruct | 10B | 2023-05-28 | not stated in card front matter | GLM-10B (card: the 10-billion-parameter General Language Model) | Download |
| WebGLM-2B | instruct | 2B | 2023-06-14 | not stated in card front matter | GLM-2B (card: the 2-billion-parameter General Language Model) | Download |
GLM (original, 2021-22) from 2023-02-28
The original blank-infilling GLM pretrained models (ACL 2022); Hub repositories created February-March 2023
| Model | Edition | Parameters | Released | Licence | Built on | Get it |
|---|---|---|---|---|---|---|
| GLM-10B | base | 10B | 2023-02-28 | not stated in card front matter | Download | |
| GLM-10B-Chinese | base | 10B | 2023-02-28 | not stated in card front matter | Download | |
| GLM-2B | base | 2B | 2023-03-01 | not stated in card front matter | Download | |
| GLM-Large-Chinese | base | not stated (large) | 2023-03-01 | not stated in card front matter | Download | |
| GLM-RoBERTa-Large | base | not stated (large) | 2023-03-01 | not stated in card front matter | Download |
Copies with the refusals removed on Hugging Face, the most downloaded counted up to five per model and not linked: GLM-4.5-Air 5, GLM-4.7-Flash 5, GLM-5.3 5, GLM-5.3-Flash 5. Other people publish copies of open-weight models with the refusals removed, usually called abliterated or uncensored. We count them and don't link them. They're unvetted (third-party uploads are where tampered files turn up), they still carry the base model's licence and its use policy, and they tend to follow instructions and call tools less reliably, which is what an agent needs most. Get the weights from the lab's own repository.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.