The team has open-sourced the world's fastest WebGPU kernels for running local AI models directly in browsers via WebGPU. The collection covers over 200 common machine learning operations that execute entirely client-side, with ongoing efforts to integrate these optimizations into frameworks like Transformers.js and ONNX Runtime Web. Detailed implementations and performance benchmarks are available at the Hugging Face kernel repository.