LFM3-Omni-2B

A single model for four modalities

LFM3-Omni-2B is the first omni-modal model in Liquid AI's LFM3 family^[1]. A single set of weights takes text, images, video and audio as input and answers in text or in speech^[1]. The model holds 2.17 billion parameters in total, and its weights were uploaded to Hugging Face on 21 October 2026^[1].

The release is framed around consolidation. Liquid AI presents the model as a replacement, on device, for the separate speech-recognition, vision-language, language-model and text-to-speech models that such a system would otherwise need^[1]. In a conventional setup each of those jobs is handled by its own network with its own weights, preprocessing and latency; here one model covers all of them.

How it is built

Four components make up the architecture: an LFM3 backbone, an LFM3 vision encoder, a causal audio encoder and the LiquidCodec audio tokenizer^[1]. The causal audio encoder handles incoming sound, and LiquidCodec turns audio output into speech, generated at 24 kHz^[1].

The context window is 32K tokens and covers input and output together, over a vocabulary of 128,000 tokens^[1]. Images and video are read at native resolution instead of being cropped or resized to a fixed size, and each 2×2 patch block is collapsed into one token in the language sequence^[1]. Packing four patches into a single token is what keeps high-resolution visual input from flooding a 32K budget. Video is sampled at two frames per second, and clips run from 4 to 384 frames by default^[1]. For video that carries sound, the model sees and hears the clip at the same time^[1], so speech, music and background noise in a recording are available to the same pass that reads the frames.

Languages and licence

The model supports 23 languages, among them Arabic, Chinese, Japanese, Korean and Hindi^[1]. It is released under the custom LFM 1.0 licence rather than an OSI-approved one^[1] — a detail worth checking before building on top of it, since a custom licence can carry terms that a standard open-source licence does not.

Where it sits

Within the LFM3 family, this is the entry that brings audio and video input and speech output together in one model^[1]. The combination — 2.17 billion parameters, native-resolution vision, simultaneous seeing and hearing of a clip, and speech as an output format — points at the kind of use where one small model has to cope with whatever a microphone or camera delivers, rather than a stack of task-specific models each loaded for its own job^[1].