Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI

Hugging Face has announced the release of @huggingface/kernels, a library designed to load and run optimized WebGPU kernels from the Hugging Face Hub to improve browser-based machine learning inference. The initial release includes 207 versioned kernels covering various machine learning architectures and workloads, such as matrix multiplications, convolutions, and attention primitives. Each kernel is published as a complete package featuring its interface, shader templates, correctness cases, and benchmark instructions. To evaluate performance across real-world hardware, Hugging Face also introduced Fleet, an in-browser testing suite that collects private performance and correctness evidence with user consent. According to Hugging Face, testing against ORT WebGPU on an Apple M4 GPU showed their kernels achieved a geometric mean speedup of 2.57x across matching test cases, with some specialized operations yielding significantly larger gains. Hugging Face states it is collaborating with the ONNX Runtime team to upstream these improvements into the broader ONNX Runtime Web ecosystem.