First run will download ~2.1 GB (q4 ONNX weights). Desktop Chrome/Edge recommended. HTTPS or localhost required for WebGPU.
This will download model files (~2.1 GB, quantized q4) to your browser cache and run inference via WebGPU. Proceed?
WebGPU works on secure contexts (HTTPS/localhost). Safari/Firefox may need experimental flags; fallback is WASM/CPU.