# LFM3-Audio-2B (slim)

> LFM3-Audio-2B is Liquid AI's open-weight 2B audio model for on-device use, taking text or speech and answering in text, speech or both, with transcription, translation, tool calling and a 32K-token context, released on 27 October 2026.

- Full: https://www.anchorterminal.com/tools/lfm3-audio-2b.md (~10,050 tokens) · this version ~1,880 tokens · JSON https://www.anchorterminal.com/tools/lfm3-audio-2b.json · canonical https://www.anchorterminal.com/tools/lfm3-audio-2b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/lfm3-audio-2b#pricing · reviews https://www.anchorterminal.com/tools/lfm3-audio-2b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #64 of 135 open-weight models · #54 in Open-weight models · not agent-ready · confidence low**


Assessment: A 2B model that listens and speaks on an 8 GB machine, transcribes and translates 19 languages and turns spoken requests into tool calls. Only llama.cpp and Ollama run it in a release. The context is 32K, voice chat is English only, Liquid hasn't answered the repository's questions, and there are no safety results or data sources.

## Facts

- Kind: Open-weight model · vendor: Liquid AI · category: Open-weight models · legal entity: Liquid AI, Inc. (the licensor named in the LFM Open License) · provenance 73/100
- Packages: huggingface `LiquidAI/LFM3-Audio-2B`, huggingface `LiquidAI/LFM3-Audio-2B-GGUF`, huggingface `LiquidAI/LFM3-Audio-2B-MLX-4bit`, ollama `lfm3-audio`
- Auth: None · pricing: Free · x402: no · licence: LFM Open License v1.0 (free commercial use below $10M annual revenue)
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/LiquidAI/LFM3-Audio-2B and Liquid's GGUF, MLX and ONNX repositories, read from the model card, the files and their metadata, the chat template, the head of the licence, Liquid's docs and blog post, Ollama's library, LM Studio's catalogue and the runtimes' own code at their latest releases on 10 November 2026. Liquid's licence terms, company pages and security.txt were read for the LFM2.5-2.6B listing on 21 October and 4 November, and security.txt again today. Nothing was downloaded or run
- Model: 2,002,332,928 parameters in the files (2B on the card): the LFM3 backbone, a causal audio encoder that works at 16 kHz and the LiquidCodec audio tokenizer, which codes 80 ms of speech as 12 tokens. Text or audio in; text, 24 kHz speech or the two interleaved out, streamed. It transcribes, speaks and translates speech, answers questions about recordings and can reason before a text reply
- Built on: The LFM3 backbone, per the card. The repository's metadata names no base model, and the licence is the same LFM Open License v1.0
- Files: Safetensors in BF16 (4.17 GB for the repository) with chat_template.jinja, config, tokenizer, a LICENSE file and 9 Python files that Transformers loads as the model's code. No pickle files. Hugging Face lists a SHA-256 for each of the 3 large files
- Licence: LFM Open License v1.0, Apache-like terms (copyright and patent grants, redistribution with notices) with a commercial-use limit. Commercial use is licensed only for legal entities under $10M in annual revenue, as read in full for the LFM2.5-2.6B listing. The LICENSE file is in the repository
- Runtimes: Released llama.cpp has the lfm3 architecture with the LFM3 audio projector and speech output (src/models/lfm3.cpp and tools/mtmd at b12242, with a realtime client example that starts this model), and Ollama's library has an lfm3-audio entry. The card names vLLM and SGLang for the BF16 weights, and neither 0.33.0 nor 0.5.23 has lfm3 code. Transformers 5.22.2 loads it only through the repository's own Python, and LM Studio's catalogue has no entry for it. The card asks for transformers 5.20.2 or later with trust_remote_code
- Context: 32K tokens of input and output per the card. config.json gives the language model 128,000 positions and Liquid's API lists the hosted model at 128,000, but the card's figure is the one counted
- Tools and reasoning: Tools go in through `tools=` in apply_chat_template and calls come back as XML, a superset of the qwen3_coder format, which Liquid's docs name as the parser for vLLM. Tools marked `defer_loading` stay out of the prompt until a search tool returns them. The card documents reasoning per request for text replies: the caller prefills the assistant turn with an open think tag to let the model reason, or a closed one for a direct answer, with a use_thinking example. Speech replies don't reason
- Hardware: No memory figure on the card. Liquid's Q4_K_M GGUF is 1.10 GB with its 0.17 GB projector file, its Q8_0 1.85 GB, its 4-bit MLX build 1.73 GB and the BF16 files 4.16 GB. Liquid also publishes an ONNX repository. It has no FP8 or NVFP4 build of this model
- Training data: The card has no training section. Liquid's blog post describes the shared LFM3 backbone's 22T tokens of pre-training in general terms and says nothing about the speech and audio data this model was trained on. No sources or cut-off
- Languages: 19 speech languages and locales, listed on the card

- Scores: Capability pending, Reach 80, Risk 40, Agent interface 85, Documentation and upkeep 52 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, Liquid's Q4_K_M GGUF is 1.10 GB with its 0.17 GB projector file, and 1.27 GB × 1.2 = 1.5 GB, which fits 8 GB, well inside the 16… · Risk, File safety 15, safetensors only, no .bin, .pt or .pth files (25). · Agent interface, Chat template 20, chat_template.jinja in the repository (9,328 characters), with roles, audio, tools, deferred tools and tool results. · Documentation and upkeep, Model card 17, Intended use (local agents, tool calling, structured outputs, fine-tuning), limitations (not recommended for agentic coding o…
- Sources: 26, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/lfm3-audio-2b.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: Discord https://discord.gg/DFU3WQeaYD; https://docs.liquid.ai/lfm/models/lfm3-audio-2b; more in the full twin

## Before you call it

1. Start it with `llama-server -hf LiquidAI/LFM3-Audio-2B-GGUF:Q4_K_M --jinja` or `ollama run lfm3-audio`
2. llama.cpp ships examples/realtime_client.py for live speech with this model
3. Prefill the assistant turn with an open think tag to let it reason before a text reply
4. The context is 32K
5. Voice chat is English only, while transcription, synthesis and translation cover 19 languages and locales
6. Liquid doesn't recommend it for agentic coding or complex multi-step reasoning

## Connect

```bash
llama-server -hf LiquidAI/LFM3-Audio-2B-GGUF:Q4_K_M --jinja
# or
ollama run lfm3-audio
```

```bash
curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)