DeepSeek open-sources its Ascend training stack, putting domestic compute under real load
DeepSeek has ported and open-sourced its full training stack for Huawei Ascend, the first stack-level validation of domestic AI compute.
ImportanceMaterialEvidenceE3 inspectableWrite-upStandard
On September 30, DeepSeek announced via its official WeChat account that it has open-sourced its training infrastructure components for Huawei's Ascend platform, matching its earlier NVIDIA releases one for one.
The release covers the TileLang compiler, the DeepGEMM matrix-multiplication library, the DeepEP cross-device communication library, plus TileKernels, FlashMLA and DeepSelect; the code is public on GitHub. The company says most operators in its V4-series model training are implemented in TileLang, and that every one of them now has a corresponding Ascend implementation.
Claims of near-hardware-limit performance come from DeepSeek itself, with no third-party tests in the announcement. The port and the public code themselves, however, show Ascend can carry its main training workload — the most direct stack-level validation of the domestic compute ecosystem so far.