DeepSeek releases six open tools for Huawei Ascend – with TileLang support for Ascend 950
The company released six open modules tailored for Huawei Ascend chips on Wednesday, including an Ascend-compatible version of the TileLang programming language, which now officially supports the Ascend 950. The release is part of a coordinated software-and-hardware push to build a Chinese AI stack outside Nvidia – but the most interesting claims remain independently unverified.
DeepSeek released six open-source modules adapted for Huawei's Ascend AI chips on September 30, 2026 – a Wednesday – in an announcement on the company's official WeChat account. The company describes the release as a move to build a new "independent and controllable" software ecosystem for graphics processors, not just a single product, according to the South China Morning Post.
The six modules mirror the open tools DeepSeek previously released for Nvidia's AI chips. This time the target is Huawei's Ascend platform, amid US export restrictions that have made Chinese companies dependent on domestic alternatives to Nvidia.
What was released – and what TileLang actually is
The most central tool released is an Ascend-compatible version of TileLang, a high-level programming language for developing high-performance kernels – the essential computational programs that make AI models run efficiently on GPUs and CPUs.
According to an update on TileLang's GitHub page, cited by SCMP, the language still lists Nvidia as its primary backend, but now officially supports Huawei's Ascend 950 accelerators with "native code generation, automatic scheduling and synchronisation". DeepSeek additionally states that the toolkit is free to download and includes compute and communication libraries, according to Mobile World Live.
The CUDA framing – the company's own claim
DeepSeek itself framed TileLang as a rival to Nvidia's CUDA, claiming the language offers "a simpler programming model", according to Quartz/Yahoo. In the announcement, as cited by The New York Times, the company presented the Huawei partnership as a potential CUDA alternative: for a successful AI ecosystem to emerge, specialised software must be "a universal, easily programmable language that can also reach the hardware's performance ceiling".
It is worth being clear about what is confirmed and what is not: the "rival to CUDA" framing is DeepSeek's own characterisation. No independent source documents how the tools actually perform on the Ascend 950 compared with CUDA on Nvidia hardware. Whether a high-level language can reach "the hardware's performance ceiling" in practice remains an open question.
It is nonetheless understandable why the software layer is the strategically interesting part. Nvidia's dominance rests not only on chips but on CUDA – a software platform that, according to a NYT estimate, is used by an estimated four million developers worldwide (an estimate, not a verified figure). By open-sourcing a competing programming language, DeepSeek is attempting to attack this ecosystem moat directly, not merely offer yet another chip.
The hardware partnership and Huawei's accelerated roadmap
The software release closely follows Huawei's own hardware offensive. DeepSeek and Huawei have collaborated on a "supernode" solution based on 128 Ascend 950 chips, optimised for both computation and communication, DeepSeek says. The release came roughly two weeks after Huawei launched its next-generation AI processors and supernode computing systems, which the company says it expects to put into operation for model training next year.
A week before DeepSeek's release, Huawei further announced that the Ascend 960DT chip will arrive in the first quarter of 2027 – three quarters earlier than originally planned – with a roadmap extending to 2029.
Nor is this the first time DeepSeek has moved away from Nvidia. Earlier this year, the company launched a preview of its V4 model adapted for Huawei's Ascend processors, and Huawei said at the time that it had coordinated with DeepSeek to ensure V4 compatibility across the Ascend product line.
The Huawei chief's market-share claim – without data
In connection with Huawei's launch, Huawei's rotating chairman Eric Xu said the company believes Ascend chips hold a larger share of China's AI chip market than Nvidia. However, he presented no data supporting the claim, Quartz/Yahoo notes – and it should therefore be read as an unsupported company statement, not verified market statistics.
What remains to be seen
Several open questions remain before the significance of the release can be assessed. There are no independent performance benchmarks comparing the tools with the CUDA ecosystem. And whether six modules amount to a usable ecosystem – with enough developers, documentation and tooling support to constitute a genuine alternative to Nvidia's established CUDA investment – is still unknown.
What DeepSeek has demonstrated is a methodology: the same open-source strategy that brought the company global attention for its models is now being applied to the software layer beneath the chip. Whether that is enough to crack an ecosystem of an estimated four million developers is a question of time and adoption – not of intent.