Liquid AI's LFM3-Omni-2B brings any-to-any AI to the edge

Liquid AI has released LFM3-Omni-2B, a compact multimodal model that aims to pack a wide range of input and output types into a footprint suited for edge hardware. At roughly 2 billion parameters, it is designed to run where cloud-scale models cannot.

The model is described as "any-to-any," meaning it can take in and produce across multiple modalities. According to the release, that includes vision and audio understanding alongside text, plus speech-to-speech capability — a notable combination for a model of this size.

Why it matters

Most multimodal systems that handle both vision and real-time speech remain large and tethered to datacenter inference. A 2B model that spans these tasks fits Liquid AI's broader focus on efficient architectures intended for on-device use. Smaller any-to-any models open the door to:

• Local voice assistants that respond without a network round-trip
• Privacy-sensitive applications where audio and images stay on the device
• Lower-cost deployment on consumer and embedded hardware

The model is published under a custom license, so teams evaluating it for production should review the terms on the model page before building on it. As with any early release, real-world performance on vision, audio and speech tasks will be the measure of how far a model this compact can stretch.