DeepSeek partners with Huawei, open-sourcing TileLang-based programming tools and libraries for Ascend chips as a CUDA alternative
Confirmed
The takeaway
On Sept 29, 2026 DeepSeek said on its official WeChat account that it is working with Huawei on programming tools for Huawei’s Ascend AI chips.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 20 October 2026
Your AI and this story
- GPT-6.2 Astra121 days after its cutoff
- Claude Opus 5.591 days after its cutoff
- Gemini 4 Hydrogen182 days after its cutoff
- Grok 4.7121 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 91 days before it.
Key facts
- Announced on DeepSeek’s official WeChat account on Sept 29, 2026 (Reuters)
- Open-sourced: TileLang support for Ascend, plus Ascend-optimized compute and communication libraries (operators, matrix computation, inter-card communication, attention, data filtering, per secondary reports)
- DeepSeek says TileLang offers ‘a simpler programming model’ than CUDA while reaching the hardware’s full performance
- DeepSeek: ‘To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language that is universal, easy to program’ and still reaches full performance
- The two companies also worked on a supernode of 128 Ascend 950 processors, optimizing how computation and chip-to-chip communication are distributed (Reuters via Techzine)
- Background: Huawei’s own CANN stack had been criticized as hard and unstable to use; TileLang is an existing open-source tile-based kernel language that DeepSeek had used for GPU kernels before
- Released repos: deepseek-ai/DeepGEMM-Ascend (created Sept 29; ~583 stars by Oct 16), a port of DeepGEMM that is fully API-compatible and supports BF16, FP8 and FP4 GEMM, MQA logits and MegaMoE on Ascend NPUs; and deepseek-ai/DeepEP-Ascend (created Sept 30; ~257 stars), described as ‘a high-performance communication library for machine learning training and inference on Huawei Ascend NPUs’ (READMEs and repo descriptions)
- The Information (Oct 19, Qianer Liu and Juro Osawa; paywalled, Techmeme summary): DeepSeek and Huawei are collaborating on future AI chip designs, and DeepSeek plans to start using the Ascend 950DT for pre-training as soon as the end of 2026. Earlier reports summarized in the press (Bloomberg) had said DeepSeek meant to use the 950DT for running models, not training, so the training role is still contested
What happened
DeepSeek announced the partnership and the open-source release on WeChat. It covers a high-level programming language (TileLang) with an Ascend backend and libraries for compute and inter-chip communication. These mirror the toolkit DeepSeek had built for Nvidia GPUs.
Why it matters
Nvidia’s lock on AI chips rests largely on CUDA. China’s leading open-model lab is now publishing a developer stack for Huawei’s chips, which tackles the software barrier as well as the hardware one. Its V4 models already run on Ascend.
Unverified: the exact names of the released repositories were not in the reports read for this run.
Sources
9 sources from 7 sites. Numbers match the chips in the text.
9 sources: 3 primary, 6 press
Primary
- GitHub: deepseek-ai/DeepGEMM-Ascendgithub.com, code
- GitHub: deepseek-ai/DeepEP-Ascendgithub.com, code
- TileLang on GitHubgithub.com, code
Press
- Reuters via Investing.com: DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidiainvesting.com, press
- US News (Reuters): DeepSeek partners with Huawei to develop chip programming toolsmoney.usnews.com, press
- SDxCentral: DeepSeek teams up with Huawei to beef up Ascend chip programming toolssdxcentral.com, press
- Techzine: DeepSeek brings AI software to Huawei’s Ascend chipstechzine.eu, press
- Reuters: DeepSeek partners with Huawei to develop chip programming toolsreuters.com, press
- The Information: How DeepSeek is shaping Huawei’s AI chips in a growing challenge to Nvidia (paywalled)theinformation.com, press
Changes
- The Information Oct 19: chip co-design and possible 950DT pre-training from end-2026 (paywalled; Techmeme summary)
- Added the DeepGEMM-Ascend and DeepEP-Ascend repos
- Added Reuters original link
- Filed