Hugging Face released 207 WebGPU kernels and a JavaScript package, @huggingface/kernels, on September 1, 2026. The package lets developers load individual GPU operations from the Hugging Face Hub and run them inside a browser.
Kernels are small computing operations that models use repeatedly. Instead of shipping an unversioned shader file, Hugging Face publishes each operation as a versioned repository with its interface, WebGPU shader templates, correctness tests and benchmark cases. Developers can see what an operation expects and test whether it produces the right output.
The release also includes Fleet, a browser-based testing tool. It measures how kernels perform on the user's GPU and can contribute correctness and performance evidence with the user's consent. That gives Hugging Face data from more hardware combinations than it could test in one lab.
The 207 kernels carry an Apache 2.0 license. The loader is available through npm under the @preview tag, but the browser must support WebGPU. Support varies by browser, operating system, GPU and driver.
These kernels provide building blocks for local model inference; they are not, by themselves, a complete model runtime. Hugging Face says higher-level browser inference tooling is still being developed.