What happened
DeepSeek said it is collaborating with Huawei to open‑source TileLang and related and communication libraries for the Ascend chip platform.
On September 30, 2026, DeepSeek announced a partnership with Huawei Technologies to develop and open‑source programming tools for Huawei’s Ascend GPU family. The core of the effort is TileLang, a high‑level language designed to simplify AI model development on Ascend hardware. DeepSeek also plans to release associated and communication libraries that sit on top of the Ascend platform.
According to DeepSeek’s statement, TileLang is intended to be "universal, easy to program, and still capable of reaching the hardware’s full performance potential," positioning it as a more developer‑friendly alternative to Nvidia’s CUDA. The company claims the language offers a simpler programming model that can improve development efficiency and reduce code complexity.
The partnership builds on earlier collaborations where DeepSeek adapted its V4‑Flash model to run on Ascend chips. In April 2026, Huawei reported that its Ascend chips were used for part of the training of DeepSeek’s V4‑Flash model, and the two firms had previously worked on model‑level optimisations. This new effort shifts focus from model adaptation to the underlying software stack that enables such models to run efficiently.
Source details: cxodigitalpulse.com ↗
Why it matters
The move provides a domestic software stack for AI developers in China, reducing dependence on Nvidia’s CUDA ecosystem and potentially accelerating adoption of Huawei’s Ascend GPUs for large‑scale model training and . By offering a high‑level language that DeepSeek claims is easier to use while retaining performance, the partnership could lower engineering effort and cost for companies building AI applications on Chinese hardware. The initiative also aligns with broader geopolitical pressures limiting access to advanced foreign semiconductor tools, making a self‑contained AI computing stack strategically important for China’s AI industry.
Providing an open‑source, high‑level language for Ascend GPUs addresses a critical gap in China’s AI ecosystem, where most existing tools are tied to Nvidia’s CUDA. By reducing the need to rewrite or re‑optimize code for non‑Nvidia hardware, TileLang could lower barriers for Chinese firms and research institutions to adopt domestic chips, supporting national goals of technological self‑sufficiency.
The announcement also comes amid export restrictions on advanced semiconductor equipment, which have limited Chinese access to cutting‑edge foreign GPUs. A home‑grown software stack mitigates some of the impact of those restrictions, allowing AI developers to continue scaling models without relying on foreign toolchains.
While DeepSeek’s claims about performance parity and ease of use are not independently verified, the open‑source nature of TileLang invites community testing and benchmarking. If the language delivers on its promises, it could influence the broader AI hardware market by demonstrating that viable alternatives to CUDA exist.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Future updates on TileLang’s performance benchmarks, adoption rates among AI developers, and any additional open‑source components that DeepSeek or Huawei may release.
Performance benchmarks comparing TileLang‑compiled workloads against CUDA on comparable tasks will be a key indicator of the language’s viability.
Adoption metrics, such as the number of open‑source contributions, GitHub stars, and downstream projects using TileLang, will reveal community acceptance.
Huawei may announce additional tooling, such as profilers or debugging utilities, that complement TileLang and further solidify the Ascend software ecosystem.
Regulatory developments affecting semiconductor export controls could either accelerate or hinder the uptake of domestically developed AI toolchains.