Gemma 4 · LiteRT-LM · WebGPU
Skim Recap

Gemma 4 runtime, from source to GPU.

By Yui Morii

Everything below is checkable against the shipped extension rather than taken on trust — the model identifier survives into the compiled bundle.

What is actually shipped

The model identifier is part of the extension source and remains visible in the compiled bundle, making the shipped runtime easy to verify.

Technical layer Implementation
Model Gemma 4 E4B Instruct · gemma-4-E4B-it-web.litertlm
Source definition src/offscreen.ts:12
Model URL https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it-web.litertlm
Compiled bundle dist/offscreen.js:1207 contains the same Gemma 4 model URL
Runtime @litert-lm/core 0.15.0 · local WASM runtime assets
Execution LiteRT-LM + WebGPU · Chrome offscreen document
Model storage Downloaded once from Hugging Face · persisted with the browser Cache API
Text processing Skipped article text is summarized on-device and is not sent to a hosted LLM API
Overview About Privacy Stack
Blog Chrome Store