Deepseek said it is making a set of programming tools open source for Huawei’s Ascend AI chips. The core of the kit is TileLang, a new language that Deepseek claims is easier to use than Nvidia’s CUDA platform. The release also includes libraries for computation and data movement, and the code will be publicly available.
The two companies tested the tools on a supernode – a cluster of 128 Ascend 950 chips – that Huawei fully supported. TileLang was originally created by researchers at Peking University and has been used by Deepseek for about a year, first on older Nvidia hardware.
Why the software matters
Software determines how much performance a chip can deliver. By opening the code, Deepseek hopes to give Chinese developers a common language that can unlock the power of domestic chips, which have struggled to match the software ecosystem built around Nvidia’s CUDA.
Why it matters
For developers in China, the new open‑source tools could make it easier to write AI applications that run efficiently on domestic hardware. That may reduce reliance on foreign software stacks and help local chip makers get more use out of their products.
