Skip to content

Help

Browsers and requirements

Which browsers can hold part of the model, which can only chat, and how to turn WebGPU on where it is off.

A device whose browser has WebGPU can run part of the model. Any other device can still join to ask questions and read answers.

Browser Status
Chrome or Edge 113+ on a laptop or desktop Holds layers. Chrome on macOS is the tested host.
Safari 26+ on macOS, iPhone and iPad Holds layers. Do not host a big model from Safari on a Mac (see below).
Chrome 121+ on Android Holds layers, within the phone cap.
Chrome on Linux WebGPU is off by default. Turn it on with two flags (below). Not tested.
Firefox No WebGPU here yet, so it joins to chat. Not tested.
Anything without WebGPU Joins as Chat only, no memory to give.

The page checks the browser when it opens. If this device can’t hold layers, the join screen says why and shows the fix under Details.

Open these two flags, set each to Enabled, then reload the room:

chrome://flags/#enable-unsafe-webgpu
chrome://flags/#enable-vulkan

WebGPU also needs a secure page: https://, or http://localhost when you run Pooled locally.

  • To run a model: a laptop or desktop. One computer is usually enough for the smallest model, which needs 4 GB (with an 8K context; 5.6 GB for 16K).
  • A phone alone can only chat. Phones give at most 0.5 to 2 GB.
  • Disk: each device caches its own layers. The smallest model is at most a 1.8 GB download.
  • Network: devices link directly, or through a relay. See Networks and work Wi-Fi.
  • The Serve API needs Node 22 or newer on your computer, and no GPU.
  • Model weights from Hugging Face, or from another device in the room.
  • The connection library (PeerJS) from cdn.jsdelivr.net. If an extension blocks it, the room can’t connect and says so.
  • A signaling server that introduces the devices. You can run your own.