Kernels go first-class
Custom AI kernels just got a security upgrade & a compatibility cheat sheet—no more guessing what runs where. Tap to see how Hugging Face did it!
What the video says
Hugging Face just unleashed a game-changing update to its Kernels project, standardizing custom kernel workflows for the entire AI ecosystem. A new hub repository type called Kernel turns custom code into first-class citizens, complete with system cards that spell out compatibility for accelerators, operating systems, and backends.
Security got a monumental upgrade—trusted publishers and code signing now block malicious kernels before they can hijack your machine. The revamped CLI splits kernel loading from building, paving the way for agentic development, reproducible builds, benchmarking via HF jobs, and backend-specific skills.
Framework support leaps forward with Torch Stable ABI and Apache TVMFFI, breaking hardware barriers and future-proofing kernels for the next 2 years. Every kernel now ships with a system card, so users instantly know what's compatible what's not, before a single line of code runs.
This isn't just an update; it's a breakthrough that turns custom kernels from niche hacks into a standardized, secure, and agent-ready powerhouse.
