adtestbench

AI news · DeepSeek · Huawei

DeepSeek open-sources its tools for Huawei’s Ascend chips

The release gives Chinese developers a free way to program Huawei accelerators without Nvidia’s CUDA.

By Alexander Bleu , 02:00 UTC

DeepSeek published a full software toolkit for Huawei’s Ascend AI accelerators on Wednesday 30 September 2026, announcing it on its WeChat account, Parameter reports. The code is open source and free to download.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a microchip standing on the horizon.
Drawn by adtestbench from DeepSeek open sources chip tools that could let Huawei replace Nvidia in China,

The centre of the release is TileLang, a high-level language for writing chip code that plays the role Nvidia’s CUDA plays on its own hardware. Around it come DeepSeek’s libraries for matrix maths, chip-to-chip communication and long-context attention, among them DeepGEMM, DeepEP and FlashMLA, Startup Fortune reports.

CUDA is the reason many labs stay on Nvidia: moving to other chips usually means rewriting code for months. DeepSeek is deploying more than 160,000 Ascend processors at a data centre in Inner Mongolia, and it and Huawei are working on a design that links 128 Ascend 950 chips, Parameter adds.

US export rules keep Nvidia’s best chips out of China. A free software layer from a leading lab makes Huawei hardware easier for Chinese labs to adopt.

For anyone using Chinese open models, the hardware under them is moving off Nvidia, and the software to run them is now public.