LFM3-Omni-2B: Liquid AI's 2.2B Any-to-Any Model, and How to Run It Locally

LFM3-Omni-2B is Liquid AI's 2.17 billion parameter any-to-any model: one checkpoint that reads text, images, video and audio, and answers in text or 24 kHz speech. Built for phones and laptops, it ships as a BF16 checkpoint and a GGUF build as small as 1.05 GB, and runs through llama.cpp, Ollama or a browser WebGPU demo.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts, licenses, hardware requirements and install commands against the primary source before recommending a setup path.

A new top-20 entrant on Hugging Face's overall trending list

LFM3-Omni-2B entered the top 20 of Hugging Face's overall trending list about ten days after Liquid AI published it on October 21, 2026. When this entry was filed, the model page showed 218 likes and 3,080 downloads. The README lists no GitHub repo for this release, so there is no separate star count to track. The most-liked linked demo is LFM3-Omni-2B-WebGPU (30 likes), which runs the model entirely in the browser over WebGPU, no server or API key involved.

Specs at a glance

What it actually is

LFM3-Omni-2B is meant to replace a stack of separate models with one: no standalone speech-to-text model, vision-language model, chat model and text-to-speech model bolted together, just one checkpoint that handles all four. You can ask a question out loud about a photo, transcribe or translate speech, hold a spoken conversation, or ask about a video and its soundtrack at the same time, since the model watches and listens to clips together rather than processing frames and audio separately.

Images and video go in at native resolution, with each 2x2 block of image patches collapsed into a single language token to keep the input compact. Video is sampled at a fixed 2 frames per second, with 4 to 384 frames by default. On the output side, a generation_mode argument switches between plain text, interleaved text and speech, or audio-only replies, with 24 kHz speech coming out of the built-in LiquidCodec tokenizer.

Running it with transformers

The BF16 checkpoint runs through a recent Hugging Face transformers install. Liquid's own README pins a minimum version, plus the audio and vision libraries the model needs to decode its inputs:

From there the model loads through the custom any-to-any class Liquid ships with the repo, not the generic text-generation class: AutoModelForMultimodalLM.from_pretrained("LiquidAI/LFM3-Omni-2B", trust_remote_code=True). For plain text replies, Liquid recommends greedy decoding. For speech or interleaved output, it sets audio_temperature to 0.4 with an audio_repetition_penalty of 1.2, values worth copying as-is before you start tuning your own.

Running the GGUF build on a laptop or phone

Liquid publishes a separate quantized repo, LFM3-Omni-2B-GGUF, built for llama.cpp. It is small enough that even the largest quant fits on modest hardware:

That starts a local web UI and an OpenAI-compatible API on port 8080. For a one-off prompt from the terminal instead, swap serve for cli:

Prefer Ollama's workflow instead? Liquid has registered the model under a short name, so there's no need to point at the GGUF repo directly:

For speech output specifically, llama.cpp's dedicated TTS tool takes the quantized model plus its projector and writes a wav file:

Try it in the browser first

If you want to see what LFM3-Omni-2B can do before installing anything, Liquid's own LFM3-Omni-2B-WebGPU Space runs the model fully client-side in a WebGPU-capable browser. Nothing you type or upload leaves your machine, which makes it a reasonable first stop for testing image or voice prompts before committing to a local install.

Benchmarks

Liquid reports LFM3-Omni-2B's omni-modal task scores rather than separate text-only benchmarks, since the point of the model is handling mixed inputs. It scores 70.09 on Daily-Omni, 47.10 on WorldSense, and 48.25 on the speech split of OmniBench. On single-modality tasks it reports a 6.36% word error rate on English speech recognition and a 3.80 UTMOSv2 score for English text-to-speech quality, a measure of how natural the generated voice sounds to automated raters.

FAQ

What is LFM3-Omni-2B used for?

Replacing a stack of separate on-device models with one: speech recognition, vision-language understanding, conversational chat and text-to-speech, all from a single 2.17B parameter checkpoint small enough for phones and laptops.

Can it run on a phone or does it need a GPU?

Both are possible. The BF16 checkpoint through transformers or SGLang wants a GPU, but the GGUF build runs on CPU or GPU through llama.cpp, with quants as small as 1.05 GB, which is well within reach of current phones and laptops.

Does it need an internet connection to run?

No, once the weights are downloaded. Both the llama.cpp and Ollama paths run fully offline, and the WebGPU Space runs the model client-side in the browser after the initial page load, without sending prompts to a server.

Is it free to use commercially?

It ships under Liquid's own LFM Open License v1.0, listed as "other" on Hugging Face rather than a standard open-source license. Read the license text in the model repo before building a commercial product on top of it.

Where to go from here

• Want a text-only model from the same family for comparison? Browse the local AI model directory to see LFM3-Omni-2B next to other edge-sized models side by side.
• New to running GGUF models generally? See the full how-to library for other local-AI setup guides covering llama.cpp and Ollama workflows.