Runs LLaMA, Mistral, and Phi models 100% in the browser via WebGPU — no server, no data leaves the device. High-performance in-browser inference engine.