What you can do
Ask the model something and it writes a reply. Pick any word of that reply and ride inside: you drive through the model’s 26 layers and watch its next word take shape, from nonsense to the answer.
Pin any of its neurons to a value of your choosing and ride again. The road forks at the layer where your pin changes the model’s mind, and Frizzle writes the new reply from there, from the same seed, beside the one the model first wrote. Every difference between the two is yours.
For a closer look, the Deck lays out all 6,912 neurons of every layer as cards you can flip through.
New here? Take the ride: one question, from Paris to Berlin.
What it needs
- A browser with WebGPU. Frizzle is built and tried in Chrome on a Mac; other browsers with WebGPU may work but are untried. If yours has none, Frizzle says so.
- A one-time download of 722 MB, the model file, kept on your device for next time.
- Nothing else. Nothing runs on a server: the model, the engine and every value you see stay in your browser.
How it is built
The engine is C++ compiled to WebAssembly, running every step of the model on your GPU. The page is React and three.js. The model is Google’s Gemma 3 1B, instruction-tuned, in its 4-bit Q4_0 build. The source is on GitHub.