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

Hugging Face's WebGPU release packages 207 GPU operations with defined interfaces, implementations and tests so browser-AI developers can load and improve them separately. These operations are building blocks used when running models. In the team's Apple M4 comparison, 809 cases with matching outputs and reliable timings showed a 2.57-fold geometric-mean speedup over the tested ONNX Runtime Web implementation. The timing covered work on the GPU itself, leaving out setup such as loading kernels. It therefore measures operation-level performance, while the accompanying Fleet tool collects device-specific results needed to understand how those implementations behave elsewhere.