LFM3-Omni-2B runs fully locally in a browser via WebGPU

End-to-end omni-modal model running in-browser shows local multimodal inference is feasible.

LFM3-Omni-2B is an end-to-end omni model accepting visual, audio or text input and producing interleaved text and audio, with no separate ASR or TTS models. Custom WebGPU kernels run it in-browser at about 280 tok/s on an M4 Max.

Name

LFM3-Omni-2B

Params

2B

Domain

general

Read it at Reddit →

Coverage: 1 sources, in order

1. 30 Oct 16:43 Reddit Read the article →

← Front page