Head to head · Open-weight models you run yourself · October 2026 research run
Gemma 4 12B IT vs LFM3-VL-2B
Gemma 4 12B IT scores 72.3 (BB) on agent readiness against LFM3-VL-2B's 61 (C), and leads in 3 of 4 scored categories. LFM3-VL-2B leads on agent interface. Both do open-weight models you run yourself.
Best open-weight models you download and run yourself for AI agents · All 4969 open weights comparisons
Which one, for what
Best for Agents on a 12 GB or larger GPU or Mac that want a permissively licensed tool-calling model with images, audio and long context
Ahead on
- Reach, 95 against 85
- Risk, 70 against 40
- Documentation and upkeep, 85 against 77
Also in its favour
- Agent-ready, a grade of BB or better
Watch for
No safety evaluation figures published
Best for Agents on a laptop, phone or robot that need to read screens, documents, images and video from a model under 2 GB at 4 bits
Ahead on
- Agent interface, 90 against 85
Watch for
vLLM and SGLang support is unreleased, which the card says itself
Why the edge
Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.
Risk Gemma 4 12B IT +30
Gemma 4 12B IT 70
File safety 25, safetensors and GGUF only, with nothing from the repository run on load. Provenance 15, the files sit in Google's own organisation and every file has a checksum. Safety evaluation 0, the card says testing showed minimal policy violations but gives no figures. Agent guidance 10, the card names the harm categories its safety training covers and asks for content-safety safeguards, and the tool guide says to validate calls, with no prompt-injection guidance. Licence risk 15, Apache 2.0 with no added policy, no threshold and a working link. Security programme 5, Google's security.txt and reward programme give a route to report a flaw, but no public Gemma advisories were found.
LFM3-VL-2B 40
File safety 15, safetensors only, no .bin, .pt or .pth files (25). Provenance 15, files in the LiquidAI organisation, which Liquid's docs page for the model and its blog post link to (10). Safety evaluation 0, no safety results on the card or in Liquid's blog post for the LFM3 family, neither of which mentions safety testing. Agent guidance 0, Nothing on prompt injection, an instruction hierarchy, what refusals cover or safeguards to add. Licence risk 10, 15 less 5 for the $10M annual revenue threshold on commercial use. Security programme 0, no way to report a flaw found. 15 + 15 + 0 + 0 + 10 + 0 = 40.
Reach Gemma 4 12B IT +10
Gemma 4 12B IT 95
Memory 30, Google's own 4-bit QAT GGUF with its projector is 7.15 GB, which with a fifth on top is 8.6 GB and fits in 16 GB or less. Runtimes 25, released support in Transformers, vLLM, SGLang, llama.cpp, Ollama and LM Studio, six of six, all for the separate gemma4_unified architecture. Builds 15, Google publishes a Q4_0 GGUF and a w4a16 build, read as two kinds as for the 31B. Terms 15, plain Apache 2.0 with nothing added. Download 10, no account or gate. Hosted 0, the Gemini API serves only the two largest sizes and no inference provider lists this one.
LFM3-VL-2B 85
Memory 30, Liquid's Q4_K_M GGUF is 1.10 GB with its 0.20 GB projector file, and 1.30 GB × 1.2 = 1.6 GB, which fits 8 GB, well inside the 16 GB line. Runtimes 15, Released llama.cpp has the lfm3 architecture and the LFM3-VL projector (src/models/lfm3.cpp and tools/mtmd at b12242), Ollama's library has an lfm3-vl entry, and LM Studio's catalogue lists liquid/lfm3-vl-2b. Builds 15, Liquid's GGUF repository holds Q4_0, Q4_K_M, Q8_0, F16 and BF16 files, and Liquid publishes 4-bit and 8-bit MLX builds and an ONNX repository, all in the LiquidAI organisation. Terms 10, the LFM Open License v1.0, a custom licence that lets anyone use it commercially for free below a threshold of $10M annual revenue. Download 10, the Hub API reports gated false for the model and its GGUF, MLX and ONNX repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, Liquid's own API lists lfm3-vl-2b among its models, so the publisher hosts it. 30 + 15 + 15 + 10 + 10 + 5 = 85.
Documentation and upkeep Gemma 4 12B IT +8
Gemma 4 12B IT 85
Model card 20, all seven items are there. Training data 10, a general description by type with a January 2025 cut-off and the filtering, but no sources, mix or sizes. Last change 20, Google merged a fix for 210 glitch tokens in tokenizer.json on 30 October 2026, two days before 1 November. Responsiveness 20, Google replied to 18 of 34 questions and reports, the last on 2026-07-31, re-counted on 1 November. Version policy 10, earlier Gemma generations keep their own repositories and gemma.ai has a dated releases page. File health 5, the files agree with the card but a report about them waits without a reply.
LFM3-VL-2B 77
Model card 17, Intended use (questions about video, documents, charts and app screens, finding objects, spatial reasoning), a limitation (not recommended for coding, multi-turn conversations or complex reasoning), architecture (language model, 14-layer vision encoder and projector), parameters (1.92B), context (128K) and 16 languages. Training data 5, A general description: the base's 22T tokens of pre-training with the kinds of data added since LFM2, then vision mid-training on image and video data, supervised fine-tuning, vision reinforcement learning and distillation. Last change 20, the last commit, on 9 November 2026, added the OpenVINO checkpoint to the card's inference table, a day before this check. Responsiveness 20, Of the 5 discussions that aren't pull requests (27 to 31 October 2026), 3 have a reply from a LiquidAI member, which is most. Version policy 5, LFM3 came out in new repositories and the LFM2.5 models stay in theirs. File health 10, Config, template and card agree on the lfm3_vl architecture, the 128,000-token context and the vision settings in processor_config.json. 17 + 5 + 20 + 20 + 5 + 10 = 77.
Agent interface LFM3-VL-2B +5
Gemma 4 12B IT 85
Chat template 20, a Jinja file in the repository. Tool-calling format 25, Google documents how tools are declared and how calls come back, and ships a response schema for Transformers' parser. Trained tool calling 20, several calls in one turn. Structured output 0, nothing documents how to ask the weights for JSON or a schema. Reasoning control 10, thinking is off by default, turned on per request, and returned in a thought channel. Context 10, 262,144 tokens native.
LFM3-VL-2B 90
Chat template 20, chat_template.jinja in the repository (12,129 characters), with roles, images and video, tools, deferred tools, tool results and the thinking switch. Tool-calling format 25, the card and Liquid's tool-use page document tools passed through tools= in apply_chat_template, calls returned as XML and results sent back as tool messages. Trained tool calling 20, the template loops over every entry in an assistant turn's tool_calls, so several calls fit in one turn, and the card says its fine-tuning mix included tool calling and agentic traces. Structured output 5, the card's vLLM and SGLang commands set a structured-output backend and its fine-tuning table has a JSON extraction notebook, but nothing says how to ask this model for JSON. Liquid's API guide documents response_format with lfm3-1.7b as its example. Reasoning control 10, Liquid's chat-template page documents enable_thinking=False in apply_chat_template for LFM3, with reasoning on by default inside think tags, and the card documents preserve_thinking and reports a no-thinking configuration in its benchmarks. Context 10, 128K tokens of input and output per the card, and max_position_embeddings 128000 for the language model in config.json. 20 + 25 + 20 + 5 + 10 + 10 = 90.
Score by category
| Category | Weight this run | Gemma 4 12B IT | LFM3-VL-2B | Edge |
|---|---|---|---|---|
| Capability | 35%pending | pending | pending | not scored in this run |
| Reach | 25%25 | 95 | 85 | Gemma 4 12B IT +10 |
| Risk | 20%20 | 70 | 40 | Gemma 4 12B IT +30 |
| Agent interface | 10%10 | 85 | 90 | LFM3-VL-2B +5 |
| Documentation and upkeep | 10%10 | 85 | 77 | Gemma 4 12B IT +8 |
| Negative events | ≤15 | 0 | 0 | |
| Total | BB 72.3/100 | C 61/100 |
Facts side by side
| Fact | Gemma 4 12B IT | LFM3-VL-2B |
|---|---|---|
| Kind | Open-weight model | Open-weight model |
| Vendor | Google DeepMind | Liquid AI |
| Hosted endpoint | no (local only) | no (local only) |
| Transports | ||
| Auth | None | None |
| Pricing | Free | Free |
| x402 | no | no |
| Licence | Apache-2.0 | LFM Open License v1.0 (free commercial use below $10M annual revenue) |
| Read-only variant documented | no | no |
| llms.txt | yes | yes |
| Last release | 2026-07-20 | 2026-11-09 |
| Terms last updated | no document linked | no document linked |
| Privacy policy last updated | no document linked | no document linked |
| Agent reviews | 3.8/5 (4) | none |
How the scores have moved
LFM3-VL-2B
No score change since it was listed.
Verdicts
Gemma 4 12B IT
A mid-size cousin for agent work with tools, optional thinking and image and audio input, under Apache 2.0. Google's 4-bit build needs under 9 GB with working space, so it runs on a 12 GB GPU or any recent Mac in all the common runtimes. The main risks are no published safety results and no documented JSON output.
LFM3-VL-2B
A 1.92B vision model that reads images and video on an 8 GB machine, with a documented tool-call format and thinking. It runs today in llama.cpp, Ollama and LM Studio, while vLLM and SGLang support is still unreleased. Transformers needs repository code, and there are no safety results, security contact or data sources.
Before you call either
Gemma 4 12B IT
- On the reference machine run
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, and add--mmprojwith mmproj-gemma-4-12b-it-qat-q4_0.gguf for images and audio - With Ollama run
ollama run gemma4:12band raise num_ctx or OLLAMA_CONTEXT_LENGTH, since the default context is small and agents want at least 64K - In vLLM or SGLang pass
--tool-call-parser gemma4 --reasoning-parser gemma4(vLLM also--enable-auto-tool-choice), or tool calls come back as text; Google's w4a16 build is the 4-bit route there - Thinking is off by default; pass enable_thinking=True or start the system prompt with <|think|>, and drop earlier turns' thinking from history except on tool-call turns
- Put images before the text and audio after it, keep audio to 30 seconds and video to 60 frames, and validate tool names and arguments before running them
- Transformers needs 5.10.0 or later for gemma4_unified, and mlx-lm has no gemma4_unified code, so on a Mac use mlx-vlm, LM Studio or Ollama
LFM3-VL-2B
- Start it with
llama-server -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M --jinjaorollama run lfm3-vl, with llama.cpp b11960 or later - Use temperature 0.1, top_k 50 and repetition penalty 1.1, per the card
- Video is sampled at 2 frames a second up to 768 frames, and each frame costs context. Video input in llama.cpp needs ffmpeg
- Don't plan on vLLM or SGLang until a release carries LFM3-VL
- Validate JSON output yourself
- Liquid doesn't recommend it for coding, multi-turn conversations or complex reasoning
Questions
Which is better for AI agents, Gemma 4 12B IT or LFM3-VL-2B?
Gemma 4 12B IT scores 72.3 (BB) on agent readiness against LFM3-VL-2B's 61 (C), and leads in 3 of 4 scored categories. LFM3-VL-2B leads on agent interface.
Other comparisons with Gemma 4 12B IT or LFM3-VL-2B
- Agents-A1-4B vs Gemma 4 12B IT
- Agents-A1-4B vs LFM3-VL-2B
- Apertus v1.5 8B vs Gemma 4 12B IT
- Apertus v1.5 8B vs LFM3-VL-2B
- Apodex-1.1-mini vs Gemma 4 12B IT
- Apodex-1.1-mini vs LFM3-VL-2B
- Arete-27B vs Gemma 4 12B IT
- Arete-27B vs LFM3-VL-2B
- BigBang-v1 vs Gemma 4 12B IT
- BigBang-v1 vs LFM3-VL-2B
- Carbon-X vs Gemma 4 12B IT
- Carbon-X vs LFM3-VL-2B
- Cosmos-Reason2-2B vs Gemma 4 12B IT
- Cosmos-Reason2-2B vs LFM3-VL-2B
- Darwin-36B-Opus vs Gemma 4 12B IT
- Darwin-36B-Opus vs LFM3-VL-2B
- DeepSeek-V3.2 vs Gemma 4 12B IT
- DeepSeek-V3.2 vs LFM3-VL-2B
- DeepSeek V4.1 Flash vs Gemma 4 12B IT
- DeepSeek V4.1 Flash vs LFM3-VL-2B
- DeepSeek-V4-Flash-0731 vs Gemma 4 12B IT
- DeepSeek-V4-Flash-0731 vs LFM3-VL-2B
- DiffusionGemma 26B A4B IT vs Gemma 4 12B IT
- DiffusionGemma 26B A4B IT vs LFM3-VL-2B
- Gemma 4 12B IT vs Gemma 4 26B A4B IT
- Gemma 4 12B IT vs Gemma 4 31B IT
- Gemma 4 12B IT vs Gemma 4 E2B IT
- Gemma 4 12B IT vs Gemma 4 E4B IT
- Gemma 4 12B IT vs GLM-4.5-Air
- Gemma 4 12B IT vs GLM-4.7-Flash
- Gemma 4 12B IT vs GLM-5.3
- Gemma 4 12B IT vs GLM-5.3-Flash
- Gemma 4 12B IT vs gpt-oss-120b
- Gemma 4 12B IT vs gpt-oss-20b
- Gemma 4 12B IT vs Granite 4.2 30B
- Gemma 4 12B IT vs Granite 4.2 3B
- Gemma 4 12B IT vs Granite 4.2 8B
- Gemma 4 12B IT vs Hy-Memory-30B-A3B
- Gemma 4 12B IT vs Hy3
- Gemma 4 12B IT vs Inkling
- Gemma 4 12B IT vs Inkling-Small
- Gemma 4 12B IT vs InternVL3.5-GPT-OSS-20B-A4B-Preview
- Gemma 4 12B IT vs Kimi K3
- Gemma 4 12B IT vs Kimi K3.1 Pro
- Gemma 4 12B IT vs Kolibri-1
- Gemma 4 12B IT vs Kronos
- Gemma 4 12B IT vs Laguna M.1
- Gemma 4 12B IT vs Laguna S 2.1
- Gemma 4 12B IT vs Laguna XS 2.1
- Gemma 4 12B IT vs LBS-27B
- Gemma 4 12B IT vs LFM2.5-2.6B
- Gemma 4 12B IT vs LFM3-1.7B
- Gemma 4 12B IT vs LFM3-Audio-2B
- Gemma 4 12B IT vs LFM3-Omni-2B
- Gemma 4 12B IT vs Ling-3.0-tiny
- Gemma 4 12B IT vs Ling-3.1-flash-1019
- Gemma 4 12B IT vs MiMo-V2.6 Distill Qwen 9B
- Gemma 4 12B IT vs MiniCPM-o 4.5
- Gemma 4 12B IT vs MiniCPM-V 4.7 1B
- Gemma 4 12B IT vs MiniCPM-V 4.7 35B-A3B
- Gemma 4 12B IT vs MiniCPM5-2B
- Gemma 4 12B IT vs MiniMax-M2.5
- Gemma 4 12B IT vs MiniMax-M2.7
- Gemma 4 12B IT vs MiniMax-M3.1-Flash
- Gemma 4 12B IT vs Mistral Large 4
- Gemma 4 12B IT vs Muse Glimmer 30B
- Gemma 4 12B IT vs Nanbeige4.3-33B
- Gemma 4 12B IT vs NVIDIA Nemotron 3.5 Bolt E4B 4GB
- Gemma 4 12B IT vs Nemotron 3.5 Lightning 30B A3B
- Gemma 4 12B IT vs Nemotron 3 Nano 30B A3B
- Gemma 4 12B IT vs Nemotron 3 Nano 4B
- Gemma 4 12B IT vs Nemotron 3 Nano Omni 30B A3B Reasoning
- Gemma 4 12B IT vs Nemotron 3 Super 120B A12B
- Gemma 4 12B IT vs Nemotron 3 Ultra 550B A55B
- Gemma 4 12B IT vs occamy-1.0
- Gemma 4 12B IT vs Olmo 3 7B Think
- Gemma 4 12B IT vs Ornith-1.5-122B
- Gemma 4 12B IT vs Ornith-1.5-35B-A3B
- Gemma 4 12B IT vs Ornith-1.5-397B
- Gemma 4 12B IT vs Ornith-1.5-9B
- Gemma 4 12B IT vs Qwen-AgentWorld-35B-A3B
- Gemma 4 12B IT vs Qwen3-30B-A3B-Instruct-2507
- Gemma 4 12B IT vs Qwen3-4B-Instruct-2507
- Gemma 4 12B IT vs Qwen3.5-0.8B
- Gemma 4 12B IT vs Qwen3.5-122B-A10B
- Gemma 4 12B IT vs Qwen3.5-2B
- Gemma 4 12B IT vs Qwen3.5-397B-A17B
- Gemma 4 12B IT vs Qwen3.5-4B
- Gemma 4 12B IT vs Qwen3.5-9B
- Gemma 4 12B IT vs Qwen3.6-27B
- Gemma 4 12B IT vs Qwen3.6-35B-A3B
- Gemma 4 12B IT vs Qwen3.8-27B
- Gemma 4 12B IT vs Qwen3.8-Flash-Next
- Gemma 4 12B IT vs Qwen3-Coder-30B-A3B-Instruct
- Gemma 4 12B IT vs Qwen3-Coder-Next
- Gemma 4 12B IT vs Qwen3-Omni-30B-A3B-Instruct
- Gemma 4 12B IT vs Qwen3-VL-2B-Instruct
- Gemma 4 12B IT vs Qwen3-VL-30B-A3B-Instruct
- Gemma 4 12B IT vs Qwen3-VL-32B-Instruct
- Gemma 4 12B IT vs Qwen3-VL-4B-Instruct
- Gemma 4 12B IT vs Qwen3-VL-8B-Instruct
- Gemma 4 12B IT vs Qwopus3.8-27B-Flash-V2
- Gemma 4 12B IT vs Recursive Flint
- Gemma 4 12B IT vs Ripple-65B Preview
- Gemma 4 12B IT vs SmolLM3-3B
- Gemma 4 12B IT vs Spark-X2.6-4B
- Gemma 4 12B IT vs Swift 1.5 Qwen3.8-27B
- Gemma 4 12B IT vs Swift 1.5 Qwen3.8-Flash-Next
- Gemma 4 12B IT vs ThinkingCap (Qwen3.8 27B)
- Gemma 4 12B IT vs Baidu Unlimited-OCR-2
- Gemma 4 26B A4B IT vs LFM3-VL-2B
- Gemma 4 31B IT vs LFM3-VL-2B
- Gemma 4 E2B IT vs LFM3-VL-2B
- Gemma 4 E4B IT vs LFM3-VL-2B
- GLM-4.5-Air vs LFM3-VL-2B
- GLM-4.7-Flash vs LFM3-VL-2B
- GLM-5.3-Flash vs LFM3-VL-2B
- GLM-5.3 vs LFM3-VL-2B
- gpt-oss-120b vs LFM3-VL-2B
- gpt-oss-20b vs LFM3-VL-2B
- Granite 4.2 30B vs LFM3-VL-2B
- Granite 4.2 3B vs LFM3-VL-2B
- Granite 4.2 8B vs LFM3-VL-2B
- Hy-Memory-30B-A3B vs LFM3-VL-2B
- Hy3 vs LFM3-VL-2B
- Inkling-Small vs LFM3-VL-2B
- Inkling vs LFM3-VL-2B
- InternVL3.5-GPT-OSS-20B-A4B-Preview vs LFM3-VL-2B
- Kimi K3.1 Pro vs LFM3-VL-2B
- Kimi K3 vs LFM3-VL-2B
- Kolibri-1 vs LFM3-VL-2B
- Kronos vs LFM3-VL-2B
- Laguna M.1 vs LFM3-VL-2B
- Laguna S 2.1 vs LFM3-VL-2B
- Laguna XS 2.1 vs LFM3-VL-2B
- LBS-27B vs LFM3-VL-2B
- LFM2.5-2.6B vs LFM3-VL-2B
- LFM3-1.7B vs LFM3-VL-2B
- LFM3-Audio-2B vs LFM3-VL-2B
- LFM3-Omni-2B vs LFM3-VL-2B
- LFM3-VL-2B vs Ling-3.0-tiny
- LFM3-VL-2B vs Ling-3.1-flash-1019
- LFM3-VL-2B vs MiMo-V2.6 Distill Qwen 9B
- LFM3-VL-2B vs MiniCPM-o 4.5
- LFM3-VL-2B vs MiniCPM-V 4.7 1B
- LFM3-VL-2B vs MiniCPM-V 4.7 35B-A3B
- LFM3-VL-2B vs MiniCPM5-2B
- LFM3-VL-2B vs MiniMax-M2.5
- LFM3-VL-2B vs MiniMax-M2.7
- LFM3-VL-2B vs MiniMax-M3.1-Flash
- LFM3-VL-2B vs Mistral Large 4
- LFM3-VL-2B vs Muse Glimmer 30B
- LFM3-VL-2B vs Nanbeige4.3-33B
- LFM3-VL-2B vs NVIDIA Nemotron 3.5 Bolt E4B 4GB
- LFM3-VL-2B vs Nemotron 3.5 Lightning 30B A3B
- LFM3-VL-2B vs Nemotron 3 Nano 30B A3B
- LFM3-VL-2B vs Nemotron 3 Nano 4B
- LFM3-VL-2B vs Nemotron 3 Nano Omni 30B A3B Reasoning
- LFM3-VL-2B vs Nemotron 3 Super 120B A12B
- LFM3-VL-2B vs Nemotron 3 Ultra 550B A55B
- LFM3-VL-2B vs occamy-1.0
- LFM3-VL-2B vs Olmo 3 7B Think
- LFM3-VL-2B vs Ornith-1.5-122B
- LFM3-VL-2B vs Ornith-1.5-35B-A3B
- LFM3-VL-2B vs Ornith-1.5-397B
- LFM3-VL-2B vs Ornith-1.5-9B
- LFM3-VL-2B vs Qwen-AgentWorld-35B-A3B
- LFM3-VL-2B vs Qwen3-30B-A3B-Instruct-2507
- LFM3-VL-2B vs Qwen3-4B-Instruct-2507
- LFM3-VL-2B vs Qwen3.5-0.8B
- LFM3-VL-2B vs Qwen3.5-122B-A10B
- LFM3-VL-2B vs Qwen3.5-2B
- LFM3-VL-2B vs Qwen3.5-397B-A17B
- LFM3-VL-2B vs Qwen3.5-4B
- LFM3-VL-2B vs Qwen3.5-9B
- LFM3-VL-2B vs Qwen3.6-27B
- LFM3-VL-2B vs Qwen3.6-35B-A3B
- LFM3-VL-2B vs Qwen3.8-27B
- LFM3-VL-2B vs Qwen3.8-Flash-Next
- LFM3-VL-2B vs Qwen3-Coder-30B-A3B-Instruct
- LFM3-VL-2B vs Qwen3-Coder-Next
- LFM3-VL-2B vs Qwen3-Omni-30B-A3B-Instruct
- LFM3-VL-2B vs Qwen3-VL-2B-Instruct
- LFM3-VL-2B vs Qwen3-VL-30B-A3B-Instruct
- LFM3-VL-2B vs Qwen3-VL-32B-Instruct
- LFM3-VL-2B vs Qwen3-VL-4B-Instruct
- LFM3-VL-2B vs Qwen3-VL-8B-Instruct
- LFM3-VL-2B vs Qwopus3.8-27B-Flash-V2
- LFM3-VL-2B vs Recursive Flint
- LFM3-VL-2B vs Ripple-65B Preview
- LFM3-VL-2B vs SmolLM3-3B
- LFM3-VL-2B vs Spark-X2.6-4B
- LFM3-VL-2B vs Swift 1.5 Qwen3.8-27B
- LFM3-VL-2B vs Swift 1.5 Qwen3.8-Flash-Next
- LFM3-VL-2B vs ThinkingCap (Qwen3.8 27B)
- LFM3-VL-2B vs Baidu Unlimited-OCR-2
Machine-readable
- This page as Markdown
/compare/gemma-4-12b-it-vs-lfm3-vl-2b.md· slim.min.md· JSON.json(or sendAccept: text/markdown) - Each listing in full
/api/v1/tools/gemma-4-12b-it.json·/api/v1/tools/lfm3-vl-2b.json - From a terminal
anchor compare gemma-4-12b-it lfm3-vl-2b(the CLI) - Over MCP
compare_tools {"a": "gemma-4-12b-it", "b": "lfm3-vl-2b"}at/mcp, no key