Tasks and inputs
LFM3-Omni-2B is listed for Any-to-any, Audio-text-to-text, Multimodal. It accepts text, image, audio, video and produces text, audio. Use the publisher’s input format and preprocessing for your chosen deployment.
LFM3-Omni-2B is Liquid AI's open any-to-any model: image, video, audio and text in; text and speech out, with multimodal tool calls on-device.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under LFM Open License v1.0; no hosted price confirmed.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Point a camera at a machine, ask out loud what the warning light means, and hear the answer. LFM3-Omni-2B is built for exactly that kind of interaction on low-power hardware: a 2.17B-parameter model that takes images, video (with or without its soundtrack), 16 kHz audio and text, and replies in text, 24 kHz speech, or both interleaved.
Liquid AI released it on 27 October 2026 as the most general member of the LFM3 family.
This is a non-reasoning model. enable_thinking=True is not supported. That keeps replies fast, which suits spoken interaction, but it also means harder visual reasoning falls to other models. Liquid's own LFM3-VL-2B reasons and scores better on vision-only work.
| Item | Detail |
|---|---|
| Weights | LiquidAI/LFM3-Omni-2B |
| Parameters | 2.17B |
| Encoders | LFM3 vision encoder; causal audio encoder; LiquidCodec tokenizer |
| Context | 32K tokens, input + output |
| Images | Native resolution, up to 2,048 tokens each |
| Video | 2 fps, 4 to 384 frames by default |
| Output modes | text, audio, interleaved via generation_mode |
| Languages | 16 for text; 19 speech locale entries |
| Licence | LFM Open License v1.0 |
For speech output, Liquid suggests audio_temperature=0.4 and audio_repetition_penalty=1.2. Text output works fine greedy.
Liquid reports Daily-Omni 70.1, the top score in its table and ahead of Qwen2.5-Omni-3B's 58.8, and RealWorldQA 64.7. DocVQA comes in at 92.5.
The weaker rows are just as instructive. OmniBench (46.3) trails Qwen2.5-Omni 7B. VoiceBench (64.7) trails Gemma 4 E4B (72.9), and internal tool calling (52.2) does too. English ASR WER of 6.4 is behind Phi-4-multimodal. In short: a strong generalist for audio-visual questions, not the best at any single channel.
Pass functions via tools= in apply_chat_template(). The model emits <tool_call> blocks; your app parses and executes them and returns results as tool messages. The useful part is that the request can be spoken and the context visual, so "order another one of these" with a photo becomes a function call.
--video-fps; video is fixed at 2 fps there.trust_remote_code=True.Test each channel alone first, then together. Measure ASR on your audio, VQA on your images, and speech quality by ear. Then run combined tasks: a spoken question about a short clip. If the combined task is no better than chaining specialist models, chain the specialists. If it is close and simpler to run, the single checkpoint wins.
A much larger open alternative is Nemotron 3 Nano Omni 30B A3B Reasoning. Liquid's d1-Omni-2B decision model is built on this one. More under any-to-any and on Liquid's provider page.
Compare the model on the work you need to complete.
LFM3-Omni-2B is listed for Any-to-any, Audio-text-to-text, Multimodal. It accepts text, image, audio, video and produces text, audio. Use the publisher’s input format and preprocessing for your chosen deployment.
Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.
The weights are downloadable under LFM Open License v1.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
google/gemma-4-E2B-it is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.
google/gemma-4-E4B-it is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.
lmstudio-community
lmstudio-community/gemma-4-E4B-it-MLX-4bit is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.
Qwen
Qwen/Qwen3-Omni-30B-A3B-Thinking is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.
Qwen
Qwen/Qwen2.5-Omni-7B is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.
Anthropic
Balanced Claude model for coding, analysis, agent workflows, and cost control