On-device · WebGPU · no server

A model in your browser.

A language model that runs on your GPU, in this tab. The weights download once and cache. Nothing you type leaves your machine. It knows what this site says and not much else, and a one-billion-parameter model will still get things wrong sometimes. That gap is the demo: compare it with the hosted Claude at Ask my portfolio.

Pick a model. It downloads once and stays cached in this browser.Llama 3.2 · 1B≈ 700 MB · the better answers; wants a real GPUQwen 2.5 · 0.5B≈ 350 MB · half the download; gets dates wrong more often

Runtime: WebLLM compiles the model to WebGPU shaders; inference is in the browser process.
Grounding: The system prompt is a 1,200-token summary generated from the site’s content files, sized for a small context window.
Why: Private, free to run, works offline after the first load. The server version at Ask my portfolio is the comparison.