New kernel standard
Custom kernels just got a major upgrade—now AI can write, optimize, and run them. Here’s what changed:
What the video says
Over 16,000 custom kernels now live on the Hugging Face Hub, each one a high-conviction bet to speed up AI models. Today, the platform turned those bets into a standard—a brand-new repository type called Kernel that makes every custom kernel a first-class citizen, complete with system cards showing which accelerators, operating systems, and backends it supports.
That visibility is a turning point. Developers can finally track trends across kernels, models, and applications in one place, But the bigger swing is security—trusted publishers and code signing now block malicious kernels before they ever reach a machine.
The command-line tools got a decisive overhaul too, splitting the kernel loader from the kernel builder so agents can take over the workflow. Reproducible builds, benchmarking through Hugging Face jobs, and backend-specific skills all baked in.
It's a bold move toward agentic development—the same CLI that loads a kernel today can build, test, and deploy it tomorrow, all while dynamically linking libs to DC++ to avoid the segfaults that used to plague older versions. With PyTorch and Apache TVM already on board, and 2 years of Torch Stable ABI support locked in, the update doesn't just standardize custom kernels; it paves the way for AI to write, optimize, and run them on its own.
