@huggingface/kernels Ships 207 WebGPU Kernels for In-Browser AI

What shipped

Hugging Face released @huggingface/kernels, a JavaScript library for loading and running optimized WebGPU kernels from the Hugging Face Hub, alongside an initial collection of 207 kernels hosted at webgpu-kernels. Also launching: Fleet, an in-browser GPU benchmarking suite that crowdsources performance and correctness evidence from real devices.

What changed

Until now, WebGPU ML inference in the browser was built on ad-hoc shaders — hard to discover, test, or compare across devices. @huggingface/kernels treats each GPU operation as a first-class, versioned package on the Hub: every kernel ships with its interface, WGSL shader templates, correctness tests, and benchmark cases in a single repository. The initial 207 kernels cover matrix multiplication, attention primitives, normalization, convolutions, quantization, and data-layout transforms used across ML architectures.

Fleet runs the kernels on your hardware in-browser and submits private evidence that helps the WebAI team find regressions, tune workgroup sizes, and optimize for real-world GPUs (not just lab hardware).

The library is Apache-2.0 licensed.

Why a builder cares

This is the foundation layer for fast, reliable browser inference. Instead of each runtime reinventing GPU primitives, builders get discoverable, benchmarked, versioned kernels they can import from the Hub. If you're building Transformers.js plugins, browser-based agents, or edge AI, this is the package that makes WebGPU performance reproducible across a fleet of unknown client GPUs.