Hugging Face Kernels

Full Guide

Hugging Face Kernels is a product or capability from Hugging Face documented by first-party sources. Hugging Face published a major Kernels redesign on July 6, 2026, including a first-class Hub repository type for compute kernels, stricter publisher controls, redesigned command-line tools, and broader framework support.

What it does

Kernel repositories expose compatibility information for accelerators, operating systems, and backend versions. The project now trusts approved publishers by default, supports reproducible builds and signing workflows, separates the kernels and kernel-builder command-line responsibilities, and adds support for the Torch Stable ABI and TVM FFI alongside its existing Torch path.

Availability and pricing

The Kernels library and builder are open-source projects, and compatible kernels can be browsed on the Hugging Face Hub. Hugging Face notes that signature verification is supported by the tooling but is not yet automatically enforced when a kernel loads; compute, Hub, and enterprise costs depend on the infrastructure and services a team uses.

Who it is for

AI engineers, inference platform teams, open-source maintainers, and researchers packaging, distributing, discovering, or benchmarking custom compute kernels.

What teams should review

Native kernels run with the Python process’s privileges, automatic signature enforcement is not yet active, and teams still need independent correctness, compatibility, security, and performance validation on their own hardware.

Official sources

• Official announcement
• Product or documentation page
• Technical documentation
• Pricing or plan information