gemma4-webgpu

A browser-local inference engine for Gemma 4 E2B. Purpose-built WGSL over raw WebGPU. Every token generated on your GPU — no server, no API key, no upload.

uninit Idle · load the model to begin.

Load the model

First load streams the Q4_K_M GGUF (~2.9 GB) from Hugging Face and CPU-dequantizes it to F16 on the GPU. Budget 70–90 s on a fast link. After that every token is generated locally; reopening the tab reloads from the browser's HTTP cache.

Requires a WebGPU browser with the shader-f16 feature.