@huggingface/kernels is an AI Code tool. JavaScript library for loading optimized WebGPU kernels for browser-based AI inference. Key features include 207+ Optimized Kernels, Hub-Based Loading, and Fleet Benchmarking Suite. Best for software developers and engineers, data scientists and analysts and content creators.
About @huggingface/kernels
Key Features
<strong>207+ Optimized Kernels.</strong> Access a collection of WebGPU kernels covering matrix operations, attention mechanisms, normalization, quantization, and other ML operations used across different AI architectures.
<strong>Hub-Based Loading.</strong> Download and run versioned kernels directly from Hugging Face Hub with simple JavaScript calls, eliminating the need to bundle shader code in your application.
<strong>Fleet Benchmarking Suite.</strong> Test kernel performance and correctness on your actual hardware using the in-browser Fleet tool, which crowdsources evidence across real-world GPUs and devices.
<strong>Complete Versioned Packages.</strong> Each kernel includes its interface, WGSL shader templates, correctness tests, benchmark cases, and usage instructions all in one versioned repository.
<strong>Browser-Native AI Inference.</strong> Run machine learning models entirely in the browser using GPU acceleration, enabling offline-capable AI applications without API calls or cloud dependencies.
<strong>Open Source and Free.</strong> Released under Apache-2.0 license with full source code available, allowing developers to use, modify, and reference implementations for custom WebGPU kernel development.
Frequently Asked Questions
@huggingface/kernels is a JavaScript library that lets developers load and run optimized WebGPU compute kernels for running AI models directly in web browsers. It provides 207+ pre-built kernels for operations like matrix multiplication, attention, and normalization that power local AI inference without needing a server.
Yes, @huggingface/kernels is completely free and open source under the Apache-2.0 license. You can use it in commercial and personal projects without licensing fees. All kernels are available on Hugging Face Hub at no cost.
@huggingface/kernels requires a browser with WebGPU support. Chrome 113+ and Edge have WebGPU enabled by default. Firefox requires enabling WebGPU in settings. Safari support varies by version. You can check compatibility by testing for 'gpu' in navigator in JavaScript.
Benchmarks show @huggingface/kernels performs 2.57x faster than ORT WebGPU on Apple M4 GPU across 809 test cases. The library stands out by providing individually versioned, testable kernels with built-in correctness validation and performance evidence from real devices through the Fleet benchmarking tool.
