New LFM3 Omni 2B WebGPU

This is a web demo Liquid AI built for its LFM3-Omni-2B model. It uses WebGPU to run the model directly in your browser, without going through a cloud server.

It’s a small omni-modal model: you can ask questions by typing or by speaking directly, and it responds with both text and voice. Besides chatting, it can also call on some small tools available in the browser—such as telling the time, checking the current location, and checking the battery level—making it a mini voice assistant that can work offline.

The 2B parameter size means it can run on an ordinary laptop, but don’t expect it to answer complex questions like a large model. Its greater value lies in showing developers clearly what an “on-device voice assistant” can currently do. It requires a computer and browser that support WebGPU; recent versions of Chrome and Edge generally work. Both the model weights and the GGUF version can be downloaded from Hugging Face.