LiquidAI's LFM3-Omni-2B runs on this device's GPU with WebGPU. It reads text, photos and speech and can answer out loud; nothing you send leaves the device. It needs Safari 26 or later on iPhone, or a recent Chrome, Edge or Safari on a computer.
Choose it here. Only the text model's part is read (3.5 GB), converted on the GPU, and kept in this browser for next time.
No file chosen
By Lafayette JH Compton · Toaster Club LLC. Model: LFM3-Omni-2B by Liquid AI, under the LFM Open License v1.0. Not affiliated with Liquid AI.
What can I help with?
Memory
What the assistant has learned about you across chats. It stays on this device and is loaded into the model's state when a chat starts, so nothing is looked up while you talk.
Instructions
How the assistant should behave in every chat: its system prompt. Leave it empty for the default, “You are a helpful personal assistant.” Your memory is added after it, and in voice mode “Respond with interleaved text and audio.” goes last so the model can speak. It stays on this device.
0 / 2000
About
LFM3 in the Browser is made by Lafayette JH Compton and Toaster Club LLC. It runs Liquid AI's LFM3-Omni-2B on this device's GPU with WebGPU.
The model's weights on this site were quantized and repacked from Liquid AI's release; they and the tokenizer are used under the LFM Open License v1.0, which limits commercial use to organizations under US$10 million in annual revenue. The app's code is under the Apache License 2.0. See the notices. LFM and Liquid AI are names of Liquid AI, Inc., used here only to say where the model comes from; this project is not affiliated with or endorsed by Liquid AI.
Pages on claude.ai can't open the camera or microphone directly. Record a short video instead: point the camera, say your question, then stop. The answer is spoken back.