Hugging Face has released 207 open-source WebGPU kernels intended to make AI models run faster inside web browsers. The accompanying JavaScript package, @huggingface/kernels, loads versioned GPU operations directly from the Hugging Face Hub and exposes a consistent interface to applications.
The collection covers building blocks such as matrix multiplication, normalization, convolutions, attention, quantization and data-layout changes. Each kernel repository includes its interface, shader templates, correctness tests, benchmark cases and usage instructions. That packaging matters because two shaders can produce the same answer yet perform very differently depending on the GPU, browser, input shape and available WebGPU features.
Hugging Face also launched Fleet, an in-browser testing suite that measures the kernels on a user's hardware. With consent, results contribute private performance and correctness evidence, helping identify failures and choose better variants across devices the company cannot test itself. The kernels use the Apache 2.0 license. Developers still need a browser, operating system, GPU and driver combination that supports WebGPU, so compatibility is not universal.