LFM3-Omni-2B-GGUF

LFM3-Omni-2B is a compact 2.2B-parameter omni-modal model from Liquid AI, released under the LFM Open License v1.0 for on-device deployment. Built on the LFM3-1.7B backbone with a 32K-token context window, it unifies vision, audio, and language processing to support ASR, TTS, voice chat, translation, video understanding, VQA, and tool calling. It accepts text, images, audio, and video, generating text, 24 kHz speech, or interleaved text and speech. Despite its small footprint, it outperforms several larger models on multimodal benchmarks, scoring 70.09 on Daily-Omni, 47.10 on WorldSense, 46.79 on AV-SpeakerBench, and 48.25 on OmniBench. It also achieves 92.47 on DocVQA, 806 on OCRBench, 89.34 on POPE, and 3.80 UTMOSv2 for English TTS quality. Available in BF16 and GGUF formats, it supports Transformers, SGLang, and llama.cpp.

Use with llama.cpp

Use the following commands with the GGUF files from prithivMLmods/LFM3-Omni-2B-GGUF. With -hf, llama.cpp can download the model and its multimodal projector automatically. Video input is not supported in llama.cpp yet.

LFM3-Omni-2B is a non-reasoning model. With llama-server, use --reasoning off.

Speech recognition

Use the system prompt Perform ASR in <language>. to specify the language.

llama-mtmd-cli \
  -m LFM3-Omni-2B.Q4_K_M.gguf \
  --mmproj LFM3-Omni-2B.mmproj-bf16.gguf \
  --audio speech.wav \
  -sys "Perform ASR in English." \
  -p "<__media__>" \
  --temp 0

Image understanding

llama-mtmd-cli \
  -m LFM3-Omni-2B.Q4_K_M.gguf \
  --mmproj LFM3-Omni-2B.mmproj-bf16.gguf \
  --image photo.png \
  -p "Describe this image." \
  --temp 0

Speech synthesis

The --tts-lang option accepts a language code (default: en), and --tts-voice accepts female (default) or male. The original model's audio sampling parameters are stored in the GGUF.

llama-tts \
  -m LFM3-Omni-2B.Q4_K_M.gguf \
  -mm LFM3-Omni-2B.mmproj-bf16.gguf \
  -p "Hello world" \
  -o out.wav

For more options, see the llama.cpp deployment guide.

Model Files

File Name	Quant Type	File Size	File Link	Description
LFM3-Omni-2B.BF16.gguf	BF16	3.47 GB	Link	Full BF16 weights. Highest quality, largest file size.
LFM3-Omni-2B.Q3_K_L.gguf	Q3_K_L	967 MB	Link	Lower quality but usable, good for low RAM availability.
LFM3-Omni-2B.Q3_K_M.gguf	Q3_K_M	915 MB	Link	Low quality.
LFM3-Omni-2B.Q4_K_M.gguf	Q4_K_M	1.1 GB	Link	Good quality, default size for most use cases, recommended.
LFM3-Omni-2B.Q4_K_S.gguf	Q4_K_S	1.06 GB	Link	Slightly lower quality with more space savings, recommended.
LFM3-Omni-2B.Q5_K_M.gguf	Q5_K_M	1.26 GB	Link	High quality, recommended.
LFM3-Omni-2B.Q5_K_S.gguf	Q5_K_S	1.23 GB	Link	High quality, recommended.
LFM3-Omni-2B.Q6_K.gguf	Q6_K	1.43 GB	Link	Very high quality, near perfect, recommended.
LFM3-Omni-2B.Q8_0.gguf	Q8_0	1.85 GB	Link	Extremely high quality, generally unneeded but max available quant.
LFM3-Omni-2B.mmproj-bf16.gguf	mmproj-bf16	950 MB	Link	Multimodal projection file in BF16 format. Used for audio/vision/language capabilities.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp