Hugging Face ships 207 open-source WebGPU kernels to run AI models directly in the browser

Hugging Face · youtube · 2026-09-04

Hugging Face has open-sourced 207 WebGPU kernels for browser-based AI on the Hub, along with the @huggingface/kernels library that loads, validates, renders and runs them. The accompanying video explains the design: the difference between ops and kernels, why WGSL, and why kernels ship as Jinja templates instead of fixed WGSL files — the browser picks supported data types and workgroup sizes to build the fastest kernel for each device. Demos include an attention mechanism using tinyBERT weights in 20 lines of JavaScript and a 1M+ cell wave animation driven by a single matmul kernel. The new Fleet platform lets users run kernel benchmarks on their own GPUs to help optimize for untested devices.

Original post →

More from coding & agent

coding & agent channel →