Gemma 4 runs in the browser at about 255 tokens per second

Gemma 4 E2B can run inside a web browser using WebGPU, reaching about 255 on an M4 Max machine. The speed came from optimized that Fable 5 helped improve before it was shut down.

The demo and the kernels are now available on for people to try. The model used is Google’s Gemma 4 E2B it qat mobile transformers model.

Key points

  • Gemma 4 E2B ran directly inside a browser with WebGPU.
  • The reported speed was about 255 on an M4 Max.
  • Fable 5 helped optimize the before shutting down.
  • The demo and kernels are available on .
  • This points to a possible way to reduce server for some .
Read original