DeepSeek open-sources Ascend chip tools with Huawei, aiming past CUDA

DeepSeek said Wednesday it partnered with Huawei to develop programming tools optimized for Huawei Ascend chips and is open-sourcing Ascend platform infrastructure, Reuters reported from DeepSeek’s WeChat post.

China’s frontier labs still write to CUDA, Nvidia’s chip-programming system, even when they train on domestic silicon. DeepSeek publishing Ascend-native libraries with Huawei is a concrete software path around Nvidia’s stack, not another chip slide deck.

On Wednesday, 30 September 2026, DeepSeek said on its official WeChat account that it partnered with Huawei to develop programming tools optimized for Huawei’s Ascend chips, and that it is open-sourcing programming infrastructure for the Ascend platform, including compute and communication libraries. Reuters reported the post the same day. The wire is datelined Beijing and Hong Kong, and the page stamps the story September 30, 2026, 3:03 a.m. UTC. Ascend is Huawei’s line of chips built for AI math. Those lines are DeepSeek’s, as Reuters carries them.

DeepSeek said Huawei provided full support in developing that programming infrastructure. The two companies jointly advanced a “supernode” solution based on 128 Ascend 950 chips, optimizing both computation and communication. A supernode, here, is a large machine that ties many chips together so they work as one. One hundred twenty-eight is the chip count DeepSeek named. Those lines are DeepSeek’s, via Reuters.

DeepSeek highlighted TileLang, a high-level open-source programming language for AI chips. High-level means a person writes closer to the job, and further from the wiring of one chip. DeepSeek said TileLang improves development efficiency and simplifies the logic of the code, and that it offers a simpler programming model than Nvidia’s CUDA. CUDA is the system most AI software uses to talk to Nvidia chips. DeepSeek said that building a new generation of independent, self-controlled GPU software starts with a high-level language that is universal, easy to program, and still able to reach the hardware’s full performance, and that TileLang was created to meet that need. A GPU, in that sentence, is the chip that does the heavy math. Those sentences are DeepSeek’s, via Reuters.

Reuters headlined the partnership as a step that reduces reliance on Nvidia, and wrote that Chinese tech firms are deepening ties as they seek alternatives to the Nvidia ecosystem. That frame is the wire’s. The WeChat lines in the story are about Ascend tools and TileLang.

The public code that sits with the announcement is on GitHub. tile-ai/tilelang-ascend describes itself as TileLang adapted for Ascend, a way to write the small programs that run on the chip. The repository was updated on 30 September 2026. Under deepseek-ai, DeepGEMM-Ascend describes itself as a matrix-multiplication library for Huawei Ascend chips, and its page dates an initial release to 30 September 2026, for Ascend 950 devices. Matrix multiplication is the core arithmetic inside a model. Those descriptions are the repositories’.

DeepEP-Ascend, also under deepseek-ai, describes itself as a communication library for training and running models on Huawei Ascend chips. Training means teaching a model. Running a model means using one that is already taught. The library includes the handoff that sends each piece of work to the right specialist inside a mixture-of-experts model, a design that uses only part of the network at a time. The repository was created on 30 September 2026. TileKernels describes itself as a library of those small programs written in TileLang. Its 30 September 2026 note says it added a Huawei Ascend path that sits beside the Nvidia path, so the same calls can run on either. Those lines are the repositories’.

FlashMLA is DeepSeek’s library of attention kernels. Attention is the step that decides which earlier text matters for the next word. A kernel, here, is the small program that runs that step on the chip. The repository’s 30 September 2026 note says it released sparse-attention kernels for the Huawei Ascend 950, for both the first pass over a prompt and the pass that writes the reply. Sparse, here, means the model looks at a chosen slice of the text, not every earlier word. DeepSelect describes itself as the selection kernel for DeepSeek’s sparse attention. Its 30 September 2026 note says it released that kernel for Huawei Ascend, and the same page says the library still supports Nvidia CUDA. Those lines are the repositories’.

The timing Reuters gives as context: the announcement came about two weeks after Huawei unveiled its next generation of AI processors and supernode systems, and said it expected its AI systems to be widely used for model training next year. That calendar, and that expectation, are Reuters’s account of Huawei.

In plain terms, DeepSeek said on Wednesday that it is opening the software for programming Huawei Ascend chips, with Huawei’s support, including a machine the two companies advanced together on 128 Ascend 950 chips. It pointed to TileLang as a simpler programming model than CUDA. The six public repositories name the language, the math, the chip-to-chip communication, the attention kernels, and the selection step. The WeChat wording in this story is Reuters’s account of the post.

The picture is a teal and magenta desk graphic listing those six repositories: tilelang-ascend, DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA, and DeepSelect. A line at the bottom names the 128-chip Ascend 950 path. It is a diagram of the stack. It is not a photograph of a machine room, and it does not print a date.