ThunderKittens
Hazy Research (Stanford)ThunderKittens is a CUDA-embedded framework of tile primitives for writing deep-learning kernels, built on the premise that kernels expressed over register and shared-memory tiles reach hand-written performance. Because it embeds in CUDA rather than replacing it, authors can drop to raw CUDA where the abstraction runs out. Stanford's Hazy Research group develops it.
Openness
5 high confidence- license
- MIT(OSI)
- source
- public
- core features withheld
- no
The LICENSE file contains the MIT text. The repository is public and the README describes no paid tier, enterprise edition or license-gated build, so the core is ungated.
- https://raw.githubusercontent.com/HazyResearch/ThunderKittens/main/LICENSE recorded 2026-09-30
LICENSE file is the verbatim MIT text, opening with "MIT License" and granting "Permission is hereby granted, free of charge, to any person obtaining a copy" of the software.
- https://raw.githubusercontent.com/HazyResearch/ThunderKittens/main/README.md recorded 2026-09-30
README closes with "This project is licensed under the terms of the MIT license." and states "ThunderKittens itself is a header-only library" that needs no installation. Searched the body for enterprise, license key, paid, proprietary and closed: none.
Adoption
2 low confidenceThunderKittens is a header library built from source with no package-registry presence, so GitHub stars are the only adoption signal available.
- https://api.github.com/repos/HazyResearch/ThunderKittens recorded 2026-08-18
Repo metadata - stargazers_count = 3,634 - for HazyResearch/ThunderKittens.
Capability
4 medium confidenceThunderKittens is a header-only tile library that also ships self-contained GEMM, attention and linear-attention kernels for Hopper and Blackwell GPUs. FastVideo vendors it as a submodule for its own kernels, which places it one step below Apache TVM.
- https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/.gitmodules recorded 2026-09-30
FastVideo .gitmodules declares a submodule "fastvideo-kernel/include/tk" at "path = fastvideo-kernel/include/tk" with "url = https://github.com/HazyResearch/ThunderKittens.git".
- https://raw.githubusercontent.com/HazyResearch/ThunderKittens/main/README.md recorded 2026-09-30
README says "We've provided a number of ThunderKittens kernels in the `kernels/` folder, which can be easily called from your PyTorch code." and lists attention "forwards / prefill and backwards kernels" with "causal and non-causal variants".
- https://raw.githubusercontent.com/HazyResearch/ThunderKittens/main/README.md recorded 2026-09-30
README says "ThunderKittens is mainly built and tested for Hopper and Blackwell GPUs." and that the 2.0 release "brings full support for Blackwell GPUs along with MXFP8 and NVFP4 precision".
Verified 2026-09-27