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

Hugging Face has released a library of over 200 optimized WebGPU kernels, which are the individual GPU operations that underpin machine learning models running in web browsers. Each kernel is published as a complete package on the Hub with documentation, correctness tests, benchmark cases, and shader code, rather than as isolated files. The release matters because browser-based AI inference speed depends fundamentally on how efficiently these low-level operations execute across different devices and browsers, making optimized kernels a foundational requirement for fast inference. The company also launched Fleet, a browser-based tool that crowdsources performance and correctness testing data from real-world GPUs to identify failures and guide further kernel improvements.