othermultimodalRuns locally2.17B (card)
LFM3-Omni-2B: Specs, License & How to Run
Liquid AI 2.17B omni — text/image/video/audio in, text+speech out. Not LFM3-1.7B or LFM2.5.
- License
- other
- Parameters
- 2.17B (card)
- HF likes
- 227 (2026-10-31)
- First seen
- 2026-10-21
TL;DR. LFM3-Omni-2B is a multimodal model from Liquid AI. Parameters: 2.17B (card).License: other. It is listed under ChatGPT, Closed LLM APIs, ElevenLabs.
| Developer | Liquid AI |
|---|---|
| Type | multimodal |
| Architecture | lfm3_omni (architectures Lfm3OmniForConditionalGeneration). Card: LFM3 backbone, LFM3 vision encoder, LFM3 causal audio encoder, LiquidCodec audio tokenizer. text_config 24 layers, hidden_size 2048, vocab 128,000. Distinct from LFM3-1.7B (text-only 1.73B / 256K) and from LFM2.5-24B-A2B (24.0B / 2.45B text MoE). Non-reasoning: enable_thinking=True is not supported. |
| Parameters | 2.17B (card) |
| Modalities | text, image, video, audio, speech |
| Languages | English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish, Danish, Greek, Finnish, Hebrew, Dutch, Swedish, Turkish |
| Context | 32K tokens (input + output) |
| License | other — LICENSE on LiquidAI/LFM3-Omni-2B is LFM Open License v1.0 (cardData.license other / license_name lfm1.0). Licensor is Liquid AI, Inc. Section 5 conditions Commercial Use on You/Legal Entity not exceeding annual revenue Threshold of USD 10,000,000. Above the threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits using the work for non-commercial or research purposes are exempt from the threshold. Same grant on the GGUF / MLX / ONNX siblings we opened. Not legal advice. |
| Hardware | Not yet verified — Card: designed for on-device use; marketing line says “2.2B parameters.” HF usedStorage on the BF16 card is 4,499,389,852 bytes (~4.19 GiB). Siblings LiquidAI/LFM3-Omni-2B-GGUF, LFM3-Omni-2B-MLX-4bit, and LFM3-Omni-2B-ONNX were ungated on 2026-10-31. Peak VRAM and tok/s are not independently measured. (not independently tested) |
| Hugging Face likes | 227 (as of 2026-10-31) |
| Model card | https://huggingface.co/LiquidAI/LFM3-Omni-2B |
| First seen | 2026-10-21 |
LFM3-Omni-2B is Liquid AI’s first omni-modal checkpoint in the LFM3 family: one model that understands text, images, video and audio and answers in text and speech. Official hosts we opened on 2026-10-31: the ungated card LiquidAI/LFM3-Omni-2B (created 2026-10-21T16:58:40Z; last modified 2026-10-28T23:16:38Z; 227 likes / 3,080 downloads), siblings LFM3-Omni-2B-GGUF (79 likes / 38,731 downloads), LFM3-Omni-2B-MLX-4bit (10 likes / 1,070 downloads), and LFM3-Omni-2B-ONNX (9 likes / 346 downloads), the LFM Open License v1.0, the LFM3 collection, and the family post Introducing LFM3. API gated was false on every card in that set.
It is not LFM3-1.7B (text-only 1.73B / 256K) and not LFM2.5-24B-A2B (24.0B / 2.45B text MoE). Sibling facts: LFM3-1.7B vs LFM3-Omni-2B. Family facts versus that earlier Liquid MoE: LFM3 vs LFM2.5-24B-A2B.
This identification page hangs off closed LLM APIs, ChatGPT alternatives, and ElevenLabs alternatives for the speech-out mode. Those mappings are editorial. We have no sourced quality comparison against ChatGPT or ElevenLabs. There is no Qwen2.5-Omni page on this site, so there is no vs URL.
The card says the checkpoint is built for on-device use and that one model replaces a separate ASR model, vision-language model, LLM, and TTS model. That sentence stays vendor. pipeline_tag is any-to-any. The card states this is a non-reasoning model (enable_thinking=True is not supported).
What we verified (2026-10-31)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/LiquidAI/LFM3-Omni-2B (not gated) | HF API |
| Created / last modified | 2026-10-21T16:58:40Z / 2026-10-28T23:16:38Z | same |
| Likes / downloads | 227 / 3,080 | same |
| License tag + file | other / lfm1.0 + LFM Open License v1.0 |
HF cardData + LICENSE |
| Pipeline tag | any-to-any | HF API |
| Params (card) | 2.17B total (marketing line also says 2.2B) | README “Model details” |
| safetensors.parameters | 2,166,928,384 | HF API |
| usedStorage (BF16) | 4,499,389,852 bytes (~4.19 GiB) | same |
| Context | 32K tokens (input + output) | README |
| Inputs | Text, images, video ± soundtrack, 16 kHz audio | README |
| Outputs | Text, 24 kHz speech, or interleaved | README |
| Vocab | 128,000 text tokens | README |
| GGUF sibling | 79 likes / 38,731 downloads | HF API |
Text languages on the card: EN, AR, ZH, FR, DE, IT, JA, KO, PT, ES, VI, TH, ID, HI, RU, PL. Speech locales add DA, EL, FI, HE, NL, SV, TR and named English locales (British / American). Video is sampled at 2 fps, 4 to 384 frames by default.
How to run (from the official README, not tested)
Copied names only. forkalt has not loaded a checkpoint.
The card lists BF16 (SGLang, Transformers — SGLang serves text output), GGUF (llama.cpp, including a realtime speech server), plus the MLX-4bit and ONNX siblings. Install line on the card: pip install "transformers>=5.20.2" torch torchvision torchcodec accelerate pillow librosa soundfile. Sampling copied: text output, greedy; speech or interleaved, audio_temperature=0.4, audio_repetition_penalty=1.2. Optional generation_mode: "text", "audio", "interleaved".
Hosted Liquid Playground is a different artifact from the weight grant.
Vendor benches (self-reported)
The Omni card prints Liquid-measured tables and says all scores were measured by Liquid AI. Copied Omni-column audio rows only: English ASR WER 6.36, multilingual ASR 6.72, English TTS UTMOSv2 3.80, VoiceBench 64.73, tool-calling internal eval 52.20, MMAR 51.40. Those are self-reported. They are not an ElevenLabs or ChatGPT ranking and not a forkalt bake-off against Qwen2.5-Omni (no page here).
License
LICENSE is LFM Open License v1.0. Licensor is named Liquid AI, Inc. Section 5: Commercial Use rights are conditioned on You or Your Legal Entity not exceeding the Threshold of USD 10,000,000 annual revenue. Commercial Use by an entity over that threshold is not licensed under this agreement. Qualified Non-Profit Organizations using the work for Non-Commercial or Research Purposes are exempt from the threshold. Not legal advice.
Related on this site: LFM3-1.7B — LFM3-1.7B vs LFM3-Omni-2B — LFM3 (family) — LFM2.5-24B-A2B — LFM3 vs LFM2.5-24B-A2B — Drama 3 (dedicated TTS, different job) — Phonon — Nemotron 3.5 Transcribe (ASR, the other direction) — MiniCPM-V 4.7. Serving: vLLM. Catalog: home.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| English ASR WER (card, ESB / Open ASR) | 6.36 (self-reported) | https://huggingface.co/LiquidAI/LFM3-Omni-2B |
| English TTS UTMOSv2 (card, Seed-TTS-eval) | 3.80 (self-reported) | https://huggingface.co/LiquidAI/LFM3-Omni-2B |
| VoiceBench (card) | 64.73 (self-reported) | https://huggingface.co/LiquidAI/LFM3-Omni-2B |
| Audio understanding MMAR (card) | 51.40 (self-reported) | https://huggingface.co/LiquidAI/LFM3-Omni-2B |
Deploy options
- local: Hugging Face weights (BF16 / Transformers / SGLang)
- local: Hugging Face GGUF (llama.cpp)
- local: Hugging Face MLX-4bit
- local: Hugging Face ONNX
- hosted: Liquid Console API
Comparisons
FAQ
Is LFM3-Omni-2B the same as LFM3-1.7B or LFM2.5-24B-A2B?
No. LFM3-Omni-2B is the omni checkpoint in Liquid AI’s LFM3 family (text/image/video/audio in; text and 24 kHz speech out; 32K). LFM3-1.7B is the text-only 1.73B / 256K lead card. LFM2.5-24B-A2B is the earlier 24.0B / 2.45B text MoE. Same LFM Open License v1.0; different job and size. See LFM3-1.7B, LFM3-1.7B vs LFM3-Omni-2B, and LFM3 vs LFM2.5-24B-A2B.
Are the weights public?
Yes. LiquidAI/LFM3-Omni-2B is ungated (created 2026-10-21T16:58:40Z; 227 likes / 3,080 downloads in the 2026-10-31 API check). Siblings LFM3-Omni-2B-GGUF (79 likes / 38,731 downloads), LFM3-Omni-2B-MLX-4bit (10 likes), and LFM3-Omni-2B-ONNX (9 likes) were also ungated that day. Collection: huggingface.co/collections/LiquidAI/lfm3.
Can I use it commercially?
The LICENSE file is LFM Open License v1.0. Section 5 conditions Commercial Use on You/Legal Entity not exceeding USD 10,000,000 annual revenue. Above that threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits for non-commercial/research purposes are exempt from the threshold. Not legal advice.
Does it replace ElevenLabs or ChatGPT?
No sourced bake-off. Speech-out is one omni mode (card TTS prompt “Perform TTS in … using a … voice”), not a hosted voice studio. Chat mapping is editorial for a local any-to-any checkpoint. Card audio rows stay self-reported. See ElevenLabs alternatives and ChatGPT alternatives.
Sources & verification
- https://huggingface.co/LiquidAI/LFM3-Omni-2B
- https://huggingface.co/LiquidAI/LFM3-Omni-2B/raw/main/LICENSE
- https://huggingface.co/LiquidAI/LFM3-Omni-2B-GGUF
- https://huggingface.co/LiquidAI/LFM3-Omni-2B-MLX-4bit
- https://huggingface.co/LiquidAI/LFM3-Omni-2B-ONNX
- https://huggingface.co/LiquidAI/LFM3-1.7B
- https://huggingface.co/collections/LiquidAI/lfm3
- https://www.liquid.ai/blog/lfm3
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.