Hugging Face announced major updates to its Kernels software on July 6, 2026, introducing a dedicated "kernel" repository type on the Hugging Face Hub alongside redesigned command-line tools and stricter security controls.
The platform introduced restricted publishing controls to mitigate risks associated with custom code running with Python process privileges. Users must now request kernel publisher status through account settings. By default, the kernels package only loads code from trusted publishers unless developers explicitly set the trust_remote_code=True parameter.
Security and Architecture Changes
To defend against compromised accounts, Hugging Face integrated code signing using Sigstore's cosign with ephemeral private keys and GitHub workflow verification. The signing functionality is built into kernel-builder and can be verified using kernels verify-signature. Hugging Face did not say when signature verification upon kernel loading will be fully enforced.
The update also split the command-line interfaces of kernels and kernel-builder. Under the revised structure, kernels functions purely as a runtime library for loading code, while kernel-builder handles compilation and scaffolding.
For build environments targeting manylinux_2_28, Hugging Face switched from static linking to dynamic linking of libstdc++ using the official toolchain. The change prevents data corruption and segmentation faults caused by global initialization collisions when kernels interact with PyTorch.
Frameworks and Hardware Compatibility
Hugging Face expanded framework compatibility beyond PyTorch by adding support for Apache TVM FFI, which allows kernels to interoperate across PyTorch, Jax, and CuPy. The update also added support for the Torch Stable ABI, allowing kernels compiled for Torch 2.9 to maintain compatibility across subsequent Torch releases for approximately two years.
Developers can verify system compatibility using the has_kernel() method or query get_kernel_variants() to review specific rejection reasons, such as CPU architecture or operating system mismatches.
Agentic Workflows and Deployment
The updated tooling includes features designed for automated AI agents that construct, compile, and benchmark custom kernels. Agents can scaffold project layouts using kernel-builder, navigate hardware toolchains via backend-specific skills, and execute benchmark suites across different hardware families using integration with HF Jobs.
To simplify environment configuration, Hugging Face released a one-click installation script alongside a Terraform setup guide for ephemeral instances. Kernels published to the Hub automatically generate system cards as front matter, detailing exposed interfaces and usage instructions.
