DeepGEMM
DeepSeekDeepGEMM is a tensor-core kernel library for FP8, FP4 and BF16 GEMMs, fused mixture-of-experts with overlapped communication, and MQA scoring, in one CUDA codebase. Kernels compile at runtime through a JIT module, so installation performs no CUDA compilation, though running it needs the CUDA Toolkit and an NVIDIA SM90 or SM100 GPU. It borrows concepts from CUTLASS and CuTe without relying heavily on their templates.
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/deepseek-ai/DeepGEMM/main/LICENSE recorded 2026-09-30
LICENSE file is the verbatim MIT text, opening with "MIT License" and "Permission is hereby granted, free of charge, to any person obtaining a copy".
- https://raw.githubusercontent.com/deepseek-ai/DeepGEMM/main/README.md recorded 2026-09-30
README states "This code repository is released under [the MIT License](LICENSE)." and the library is built from the repository with "./develop.sh" and "./install.sh". Searched the body for enterprise, license key, paid, commercial and proprietary: none.
Adoption
2 low confidenceDeepGEMM is built from source rather than published to a package registry, so GitHub stars are the only adoption signal available. A star is not a use.
- https://api.github.com/repos/deepseek-ai/DeepGEMM recorded 2026-08-18
Repo metadata - stargazers_count = 7,700 - for deepseek-ai/DeepGEMM.
Capability
3 medium confidenceDeepGEMM covers several GEMM and attention-adjacent primitive families - FP8, FP4 and BF16 GEMMs, fused MoE with overlapped communication, and MQA scoring - rather than a single kernel. It sits two steps below Apache TVM because no project is recorded depending on it, which is where the evidence stands, not a proven ceiling.
- https://raw.githubusercontent.com/deepseek-ai/DeepGEMM/main/README.md recorded 2026-09-30
README describes "a unified, high-performance tensor core kernel library" covering "GEMMs (FP8, FP4, BF16), fused MoE with overlapped communication (Mega MoE), MQA scoring for the lightning indexer, HyperConnection (HC)" and notes "All kernels are compiled at runtime through DeepJIT".
- https://raw.githubusercontent.com/deepseek-ai/DeepGEMM/main/README.md recorded 2026-09-30
README lists requirements "CUDA Toolkit 12.9 or higher" and "CUTLASS 4.0 or higher (could be cloned by Git submodule)" and says "requiring no CUDA compilation during installation".
Verified 2026-09-27