DeepSeek has publicly released the software toolkit it built with Huawei Technologies Co for programming the Ascend line of AI accelerators. The company announced the release on its WeChat account on Wednesday, making the toolkit open-source and free to download.

The release includes TileLang, a software platform positioned as China's counterpart to Nvidia's CUDA, the platform that currently serves as the global standard for AI development. A mature programming stack matters because developers gravitate toward chips with the most accessible tooling — a factor that has reinforced Nvidia's dominance in AI accelerators.

"A sophisticated programming language with a simple coding process is essential for building an effective AI ecosystem," DeepSeek said in its WeChat post. The company added that it worked with Huawei to optimise performance on Ascend 950 chips.

Huawei provided "unreserved and vigorous support" during the research and development of the software, DeepSeek said, and the two companies will collaborate on further innovations.

Ties run deep

The open-sourcing reflects a tightening alignment between two of China's most prominent AI players, part of the small group of firms advancing Beijing's objective of substituting American software and semiconductors with domestic alternatives. DeepSeek builds open-weight models intended to compete with offerings from Anthropic and OpenAI, while Huawei leads China's effort to manufacture the accelerators that data centres require.

The two companies' collaboration already has substantial scale. DeepSeek plans to deploy at least 160,000 of Huawei's most powerful accelerators at a data centre it is building in Inner Mongolia, according to a Bloomberg News report from August. That facility would rank among the largest AI compute installations using Chinese-made chips.

Deployments of that magnitude carry significance for Huawei, which has invested heavily in the Ascend line as a response to US export controls that restrict Nvidia's most advanced accelerators from reaching Chinese customers. Widespread adoption of a viable CUDA alternative would remove one of the last practical barriers preventing large-scale migration to domestic hardware.

DeepSeek's embrace of Huawei silicon stands out because AI model training worldwide still depends heavily on Nvidia semiconductors. Most Chinese developers have continued using Nvidia hardware where they can obtain it, treating domestic accelerators as a secondary option.

Fundraising milestone

The software release coincides with DeepSeek's push toward a public listing. The start-up is finalising a funding round at a valuation close to 500 billion yuan (US$74 billion / RM304 billion), ahead of an initial public offering that could arrive as early as this year.

DeepSeek rarely discusses its internal development work, including how it integrates AI chips into its infrastructure. The decision to publish the Ascend tooling therefore marks an unusual level of disclosure for the company and signals its intent to encourage broader adoption of Huawei's ecosystem.

By releasing TileLang and supporting tools at no cost, DeepSeek gives Chinese developers a free path to writing software for domestic accelerators. The question now is whether the tooling can match CUDA's depth and documentation, which has accumulated across nearly two decades of Nvidia investment.

The stakes extend beyond DeepSeek's own operations. If Huawei's accelerators gain traction among the developers who currently default to CUDA, the balance of the AI compute market in China — the world's largest semiconductor market by demand — could shift materially.

What comes next

DeepSeek and Huawei say further joint innovations are planned, though neither company has published a roadmap. The immediate measure of success will be uptake: whether developers outside DeepSeek download TileLang, build with it, and file the fixes and extensions that sustain an open-source community.

For now, the release establishes a concrete artifact of the partnership — working, downloadable code aimed at the software layer that has protected Nvidia's position more effectively than any hardware specification.