Chinese AI firm DeepSeek partnered with Huawei on September 30 to build programming tools for Huawei’s Ascend AI chips. The collaboration targets China’s growing need for Nvidia alternatives, using open-source software and optimized libraries for Ascend processors.
DeepSeek announced the partnership through WeChat and said Huawei supported development of the Ascend programming infrastructure. The companies also advanced a 128-chip Ascend 950 supernode, optimizing computing and communication for large-scale AI workloads.
The open-source release includes TileLang, a high-level programming language designed for AI accelerator development. DeepSeek said TileLang can simplify code logic while improving development efficiency across different chip platforms.
DeepSeek also released Ascend versions of several core software components used in AI workloads. These include DeepGEMM for matrix operations, DeepEP for device communication, TileKernels for vector computing, FlashMLA for sparse attention, and DeepSelect for data filtering.
The software effort creates a broader development layer around Huawei’s Ascend hardware instead of relying solely on chip-level tools. DeepSeek said every TileLang operator used in its V4-series training now has a high-performance Ascend implementation.
The same Python interfaces can support Nvidia GPUs and Huawei NPUs, with the backend selected automatically. Such compatibility could lower development friction for teams moving workloads between different AI accelerator environments.
DeepSeek also positioned TileLang as an alternative programming approach to Nvidia’s CUDA software model. “TileLang was created precisely to meet this need,” DeepSeek said, referring to a simpler programming model for AI chips.
The partnership arrived shortly after Huawei introduced its next-generation AI processors and supernode systems. Huawei expects these systems to gain wider use for model training next year.
The development adds a software layer to China’s broader push for domestic AI computing infrastructure. Open-source tools could help developers build applications around Ascend hardware while reducing dependence on Nvidia’s mature software ecosystem.
The move also gives Huawei’s AI chips a wider developer framework for large-scale model training. It strengthens the software connection between domestic accelerators and DeepSeek’s growing AI model ecosystem.
Also Read: NVIDIA Introduces OpenShell, Sentry to Strengthen AI Agent Safety