DeepSeek has released an open-source software stack designed for Huawei Ascend AI accelerators. The suite includes TileLang, a programming language intended as an alternative to Nvidia's CUDA, alongside libraries such as DeepGEMM, FlashMLA, TileKernel, DeepSelect, and DeepEP. These tools cover various stages of AI development, including matrix multiplication and inter-chip communication.
The software is optimized for Huawei Ascend 950 chips and supports configurations linking 128 accelerators into a single supernode. The toolkit is available for free download to encourage adoption among developers. Huawei provided support during the development of these tools, which aim to reduce reliance on the Nvidia ecosystem.
DeepSeek plans to build a data center in Inner Mongolia that will house at least 160 000 Huawei Ascend accelerators. This release follows the company's April 2026 effort to adapt its V4 AI model for Huawei hardware. The new toolkit also includes benchmarking utilities to assist developers in optimizing code for the Ascend architecture.
