Benchgen
Models/liquid-ai/

LFM3-Omni-2B

DraftPublic

Model Details

LFM3-Omni-2B

Organization Context License Modality Released

Quick answer: LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family (released Oct 27, 2026): a 2.17B-parameter end-to-end model that takes text, images, video (with or without soundtrack), and 16 kHz audio as input and answers in text, 24 kHz speech, or interleaved text and speech. It is designed to replace a separate ASR, vision-language model, LLM, and TTS model on a single device, with a 32K-token context window.

At a Glance

Where LFM3-Omni-2B leads

  • Beats same-size and larger comparison models (Qwen2.5-Omni 3B/7B, Gemma 4 E2B/E4B, Phi-4 multimodal) on video+audio understanding: 70.09% Daily-Omni, 47.10% WorldSense, 46.79% AV-SpeakerBench
  • Strong spoken visual QA: 67.59% EmbSpatial, 83.78% POPE (image+audio)
  • English TTS quality: 3.80 UTMOSv2 — ahead of all compared models that report a score
  • Single end-to-end model replaces separate ASR/VLM/LLM/TTS pipeline for on-device deployment

Where it lags

  • OmniBench overall (46.32%) trails both Qwen2.5-Omni variants (47.11% / 49.65%)
  • Tool calling (internal eval, 52.20%) trails Gemma 4 E4B (72.40%)
  • Non-reasoning model — no "thinking" mode support (enable_thinking is not supported)

Best for: on-device voice/vision assistants needing a single small model for speech recognition, image/video Q&A, and spoken responses — e.g., voice chat with visual grounding on phones or laptops.

What LFM3-Omni-2B Is

LFM3-Omni-2B is Liquid AI's first fully omni-modal release in the LFM3 family, pairing the LFM3-1.7B text backbone with an LFM3 vision encoder, a causal audio encoder, and the LiquidCodec audio tokenizer. At 2.17B total parameters, it processes text, images, video (sampled at 2 fps, 4-384 frames, with or without its soundtrack), and 16 kHz audio, and can answer in text, 24 kHz speech, or interleaved text-and-speech depending on the system prompt and generation mode. Liquid frames it as a single model that can replace a separate ASR model, vision-language model, LLM, and TTS model in an on-device pipeline — ASR, TTS, translation, and audio/image/video understanding are all exposed as system-prompt-selected modes on the same checkpoint.

It is explicitly a non-reasoning model (no "thinking" mode), and ships GGUF support for a realtime local speech server via llama.cpp, positioning it for edge voice-assistant deployments rather than complex multi-step agentic work.

Specifications

FieldValue
OrganizationLiquid AI
Parameters2.17B total
Context window32,768 tokens (input + output)
ArchitectureLFM3 backbone + LFM3 vision encoder + LFM3 causal audio encoder + LiquidCodec audio tokenizer
LicenseLFM Open License v1.0 (lfm1.0)
Release dateOctober 27, 2026
ModalityOmni — text, image, video, audio in; text/speech out (16 text languages; 19 speech languages/locales)

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face in BF16 (SGLang/Transformers) and GGUF (llama.cpp, including a realtime speech server) formats.

Public Benchmark Scores

BenchmarkScoreSourceDate
POPE (image+audio)83.78LFM3-Omni-2B model card2026-10
RealWorldQA (spoken)64.68LFM3-Omni-2B model card2026-10
MMStar (spoken)51.79LFM3-Omni-2B model card2026-10
ERQA (spoken)39.43LFM3-Omni-2B model card2026-10
VoiceBench64.73LFM3-Omni-2B model card2026-10
OmniBench (overall)46.32LFM3-Omni-2B model card2026-10

Scores above are reported by Liquid AI and shown for context. They are not Benchgen measurements. All scores on a 0-100 scale. See the model card for full evaluation methodology.

Excluded from the table above (bucket-C, documented per skill procedure):

  • Daily-Omni (70.09) — now tracked; see the Daily-Omni benchmark page for the full leaderboard.
  • WorldSense (47.10) — now tracked; see the WorldSense benchmark page for the full leaderboard.
  • AV-SpeakerBench (46.79) — now tracked; see the AV-SpeakerBench benchmark page for the full leaderboard.
  • EmbSpatial (67.59), RoboSpatial — spatial/embodied reasoning evals; no local entity exists.
  • English ASR (WER 6.36), Multilingual ASR (WER/CER 6.72), English TTS (UTMOSv2 3.80), Tool calling (internal eval, 52.20) — speech-pipeline and internal evals without a confirmed local benchmark entity.
  • Audio understanding / MMAR (51.40%) — now tracked; see the MMAR benchmark page for the full leaderboard.
  • Vision-only table (MMStar 55.67, MMMU 42.44, MMBench 73.68, DocVQA 92.47, etc.) — these duplicate the sibling LFM3-VL-2B release's vision benchmark table with different (image-only, non-spoken) numbers; omitted here to avoid conflating the omni model's spoken-QA scores with a separate vision-only variant's scores.

Frequently Asked Questions

What is LFM3-Omni-2B? LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family: a 2.17B-parameter model that understands text, images, video, and audio, and answers in text and/or speech, designed to run on phones and laptops.
What is LFM3-Omni-2B's context window? 32,768 tokens for combined input and output.
How much does LFM3-Omni-2B cost? It's released under an open-weight license — free to download; you pay only your own hosting cost.
Is LFM3-Omni-2B open source? Yes — released under Liquid AI's LFM Open License v1.0 (lfm1.0).
What is LFM3-Omni-2B's knowledge cutoff? Not disclosed by Liquid AI at release.

Specs and scores sourced from Liquid AI's official Hugging Face model card (huggingface.co/LiquidAI/LFM3-Omni-2B) and release blog post (liquid.ai/blog/lfm3). Last updated 2026-11-04.