Hugging Face announced the release of @huggingface/kernels, a new open-source library containing 207 optimized WebGPU kernels designed to accelerate machine learning inference directly in web browsers. The kernels, published on Hugging Face Hub under Apache-2.0 license, cover fundamental GPU operations including matrix multiplications, convolutions, attention primitives, and quantization—the building blocks that power modern neural networks when executed locally in the browser. Each kernel is published as a versioned, self-contained package that includes its complete specification, WGSL shader implementations, correctness test cases, and benchmark data. This modular approach enables developers to discover, audit, and integrate individual operations into their own browser-based AI runtimes without relying on centralized shader files or undocumented implementations. The library includes a JavaScript loader that handles downloading, preparing, and executing kernels directly from the Hugging Face Hub. The company also launched Fleet, an in-browser GPU benchmarking and testing suite that allows developers to evaluate kernel performance on their own hardware while optionally contributing anonymized performance and correctness data back to Hugging Face. This crowdsourced telemetry approach aims to help identify performance bottlenecks and compatibility issues across diverse GPUs and browsers, ultimately improving kernel optimization and browser-based AI performance at scale.