# LFM3-Audio-2B

> LFM3-Audio-2B is a compact, fully featured audio model that runs locally on laptops and edge devices. It takes text or audio as input and responds with text, speech, or both.

- **Organisation:** LiquidAI
- **Category:** Omni / Any-to-Any
- **Parameters:** 2B
- **Context length:** 128,000 tokens
- **Licence:** lfm1.0 (Commercial use with conditions)
- **Released:** 2026-09-08
- **Hugging Face:** 18.2K downloads (30 days), 273 likes
- **Ollama:** `ollama run lfm3-audio:2b`

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 1.66 |
| Q3_K_M | 1.97 |
| Q4_K_M | 2.2 |
| Q5_K_M | 2.4 |
| Q6_K | 2.63 |
| Q8_0 | 3.09 |
| F16 | 4.93 |

## Runs with

- [llama.cpp](https://selfhost.directory/project/llamacpp)
- [LM Studio](https://lmstudio.ai)
- [MLX](https://github.com/ml-explore/mlx-lm)
- [Ollama](https://selfhost.directory/project/ollama)
- [Transformers](https://github.com/huggingface/transformers)
- [Jan](https://jan.ai) (community)
- [KoboldCpp](https://selfhost.directory/project/koboldcpp) (community)
- [LocalAI](https://selfhost.directory/project/localai) (community)
- [ONNX Runtime](https://onnxruntime.ai) (community)

## About

LFM3-Audio-2B is a compact, fully featured audio model that runs locally on laptops and edge devices. It takes text or audio as input and responds with text, speech, or both.

• Voice chat: ask questions aloud and receive streaming replies in text and speech. For text-only responses, the model can reason before answering.
• Multilingual speech: transcribe, synthesize, and translate speech across 18 languages.
• Audio understanding and reasoning: interpret recordings and answer questions about the speech, sounds, and events they contain.
• Tool calling: turn spoken requests into tool calls to control apps and devices, with text or speech alongside the calls.

LFM3-Audio-2B introduces an in-house audio encoder and LiquidCodec audio tokenizer, improving speech quality over LFM2.5-Audio. Find more information about the LFM3 family in our release blog post.

The LFM3 family includes base, text, vision, audio, omni, and decision models.

LFM3-Audio-2B is designed for local agents, tool calling, structured outputs, and domain-specific fine-tuning. Use retrieval or search for knowledge-heavy tasks. It is not recommended for agentic coding or complex multi-step reasoning due to its small size.

Choose a checkpoint variant for your inference framework and hardware.

## Alternatives

- [LiquidAI/LFM3-Omni-2B](https://selfhost.directory/models/liquidai/lfm3-omni-2b) (2.2B, lfm1.0, ≈2.16 GB at Q4): LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family: a single end-to-end model that understands text, images, video and audio, and answers in text and speech. It is built for on-device use: one model replaces a separate ASR model, vision-language model, LLM and TTS model.
- [LiquidAI/d1-Omni-2B](https://selfhost.directory/models/liquidai/d1-omni-2b) (2.2B, lfm1.0, ≈2.16 GB at Q4): d1-Omni-2B is a decision model for text, pictures, and audio, built on LFM3-Omni-2B. It returns calibrated answers in one forward pass, without generating tokens.
- [google/gemma-4-E2B-it](https://selfhost.directory/models/google/gemma-4-e2b-it) (5.1B, Apache 2.0, ≈3.81 GB at Q4): Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants.
- [google/gemma-4-E4B-it](https://selfhost.directory/models/google/gemma-4-e4b-it) (8B, Apache 2.0, ≈5.9 GB at Q4): Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants.
- [google/gemma-4-E2B](https://selfhost.directory/models/google/gemma-4-e2b) (5.1B, Apache 2.0, ≈3.81 GB at Q4): Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants.
- [deepseek-ai/Janus-1.3B](https://selfhost.directory/models/deepseek-ai/janus-1.3b) (2.1B, MIT, ≈3.24 GB at Q4): Janus is a novel autoregressive framework that unifies multimodal understanding and generation. It addresses the limitations of previous approaches by decoupling visual encoding into separate pathways, while still utilizing a single, unified transformer architecture for processing.

More: https://selfhost.directory/models/liquidai/lfm3-audio-2b/alternatives

## Runs well on

- [GeForce RTX 4060](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/rtx-4060): F16, est. 37 tok/s
- [Apple M4 (16 GB)](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/apple-m4-16gb): F16, est. 16 tok/s
- [Radeon RX 6700 XT](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/rx-6700-xt): F16, est. 53 tok/s
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/rtx-3060): F16, est. 49 tok/s
- [Arc B580](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/arc-b580): F16, est. 63 tok/s
- [Tesla T4](https://selfhost.directory/models/liquidai/lfm3-audio-2b/on/tesla-t4): F16, est. 44 tok/s

## Similar size

- [LiquidAI/LFM3-Omni-2B](https://selfhost.directory/models/liquidai/lfm3-omni-2b) (2.2B)
- [LiquidAI/d1-Omni-2B](https://selfhost.directory/models/liquidai/d1-omni-2b) (2.2B)
- [deepseek-ai/Janus-1.3B](https://selfhost.directory/models/deepseek-ai/janus-1.3b) (2.1B)
- [deepseek-ai/JanusFlow-1.3B](https://selfhost.directory/models/deepseek-ai/janusflow-1.3b) (2B)
- [Skywork/Skywork-UniPic-1.5B](https://selfhost.directory/models/skywork/skywork-unipic-1.5b) (1.5B)
- [lmms-lab/Aero-1-Audio](https://selfhost.directory/models/lmms-lab/aero-1-audio) (2.4B)

## FAQ

**How much VRAM does LFM3-Audio-2B need?**
About 2.2 GB at Q4_K_M with an 8K context. Q8_0 needs about 3.1 GB and full precision about 4.9 GB. Longer contexts need more memory for the KV cache.

**Can LFM3-Audio-2B run on the RTX 4060?**
Yes. It fits at FP16 / BF16 (about 4.4 GB), at an estimated 37 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.

**Can I use LFM3-Audio-2B commercially?**
Yes, with conditions. It is released under lfm1.0. Custom licence: read the model card before using it commercially.

**What is the context length of LFM3-Audio-2B?**
128,000 tokens (125K). The full window needs far more memory than an 8K context.

**How do I run LFM3-Audio-2B with Ollama or llama.cpp?**
With Ollama: ollama run lfm3-audio:2b. With llama.cpp: llama-server -hf LiquidAI/LFM3-Audio-2B-GGUF:Q4_K_M, which starts an OpenAI-compatible server.

## Links

- huggingface: https://huggingface.co/LiquidAI/LFM3-Audio-2B
- homepage: https://www.liquid.ai/blog/lfm3
- ollama: https://ollama.com/library/lfm3-audio

---
Source: https://selfhost.directory/models/liquidai/lfm3-audio-2b