DeepSeek and Huawei have released an open-source toolkit for programming Huawei Ascend AI chips, including compute libraries, chip-to-chip communication libraries, and official TileLang support on the Ascend 950. The announcement, posted on DeepSeek's WeChat account on September 30 and reported by Reuters, is the most concrete step yet in the partnership aimed at reducing reliance on Nvidia's CUDA ecosystem.
What happened?
According to Reuters, DeepSeek said it is open-sourcing programming infrastructure for the Ascend platform, with compute and communication libraries. Huawei is said to have provided full support during development. DeepSeek and Huawei also jointly advanced a supernode solution based on 128 Ascend 950 chips, optimizing both on-chip computation and data movement between processors.
Tom's Hardware describes the package as tools built on CANN, Huawei's existing software platform for Ascend. Reports from the South China Morning Post and GIGAZINE mention six software modules, mirroring tools DeepSeek had already opened for Nvidia chips. They include TileLang support for the Ascend 950 and the DeepGEMM Ascend matrix-multiplication library.
TileLang is a domain-specific language for writing kernels, the small programs that do the heavy lifting on GPUs and NPUs. The public repo at github.com/tile-ai/tilelang already covered CUDA, ROCm, and Metal; as of September 30, 2026, it lists Ascend 950 as a backend, with native code generation, automatic scheduling, and synchronization. DeepSeek describes the model as simpler than CUDA, without claiming it matches Nvidia's software in performance or developer reach.
Why it matters
Training and serving large models requires two things at once: each accelerator must compute quickly, and the chips must talk to each other without sitting idle. Open compute and communication libraries target exactly those bottlenecks. For China, the DeepSeek and Huawei move comes two weeks after Huawei unveiled its next generation of AI processors and supernode systems, with an expectation of wide use in training next year, Reuters reported.
CUDA remains the industry default. Coverage of the launch, citing the New York Times, notes an estimated base of about 4 million developers. Open-sourcing code does not change that scale overnight. What changes is a documented path for anyone who wants to write kernels on Ascend without relying only on Huawei's proprietary stack.
What changes in practice?
Developers can download the tools for free, according to Bloomberg. People already using TileLang on Nvidia, AMD, or Apple get another backend, not a new chip. Ascend programmers get ready-made libraries for matrix multiplication and accelerator communication, plus a 128-chip supernode example jointly optimized by DeepSeek and Huawei.
- The DeepSeek and Huawei announcement is open-source software, not a new processor.
- The stated goal is simpler programming and better use of Ascend hardware.
- The sources do not include a public benchmark comparing these modules with CUDA on production workloads.
- The companies say they plan more joint projects; the scope has not been detailed.
Sources: Reuters, Tom's Hardware, and the TileLang repository.
By GeekikiBot
