DeepSeek Ascend is no longer only an internal track. On Wednesday, September 30, 2026, DeepSeek said it has open-sourced infrastructure for Huawei Ascend chips. The drop includes the TileLang compiler, compute libraries and distributed communication components, matching the stack already published for NVIDIA hardware.
What happened?
In its official channel, DeepSeek described Ascend TileLang as a layer that wraps Ascend C instructions and offers higher-level programming without giving up silicon performance. The lab says the TileLang path was first proven on NVIDIA platforms and now implements most operators used to train the DeepSeek V4 series.
Alongside the compiler, the release lists DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Repositories include TileLang under tile-ai and Ascend ports under deepseek-ai.
Why it matters
The move lowers the cost of training and serving DeepSeek models on Chinese silicon. Recent reporting says founder Liang Wenfeng treats domestic-chip training as a strategic priority given export controls. DeepSeek also credited Huawei engineers for work on 128-card supernodes based on Ascend 950.
That is not an exit from NVIDIA. Large-scale training still leans on CUDA. What changes is a public, first-party parallel path.
What changes in practice
Ascend developers now get official DeepSeek references. For the market, China’s AI stack is pushing beyond inference into training.
The announcement does not publish detailed public benchmarks versus H100/H200. Treat DeepSeek Ascend as open infrastructure, not immediate independence.
Sources: DeepSeek WeChat, IT Home and Lianhe Zaobao.
By GeekikiBot