DeepSeek releases six open-source tools for Huawei's Ascend chips
Chinese AI firm DeepSeek has released open-source infrastructure for Huawei's Ascend platform, including compute and communication libraries, developed in partnership with Huawei. The goal is not merely to get by without Nvidia's chips, but to build an alternative to the CUDA software ecosystem itself — though so far without independent performance figures.
DeepSeek announced on 30 September 2026, via its official WeChat account, that it is open-sourcing programming infrastructure for the Ascend platform, according to Reuters coverage carried by Yahoo Finance/GuruFocus and Deccan Chronicle. The company describes it as a partnership: DeepSeek says Huawei provided "full support" during the development of the infrastructure, and the two companies say they have jointly advanced a "supernode" solution based on 128 Ascend 950 chips that optimises both computation and communication. The announcement came roughly two weeks after Huawei unveiled its next-generation AI processors and supernode computing systems, stating that the systems are expected to see wider use for AI model training next year.
What was actually released
According to Dataquest, the release comprises six open-source components covering programming, computation, communication, attention, kernels and data processing:
- TileLang — the high-level language that forms the foundation
- DeepGEMM — the compute layer
- DeepEP — the communication layer
- TileKernels — the kernels
- FlashMLA — the attention mechanism
- DeepSelect — the data processing
It is worth noting that only Dataquest gives the full list of six components. The Reuters-sourced articles mention the compute and communication libraries and TileLang, but not all six. The complete list is therefore not independently confirmed.
DeepSeek claims that the Ascend implementations correspond to components previously available for Nvidia hardware. This is the company's own assertion, so far without independent verification.
Why the software is the strategically important target
What is new in this release is not models ported to Ascend chips, but the software stack beneath the models. Nvidia's dominant position rests to a significant degree on CUDA — a mature software ecosystem that makes its GPUs practical to program for. That is the moat DeepSeek is now attacking.
Central to the effort is TileLang, an open-source high-level language for AI chips. "To build a new generation of independent, self-governed GPU software ecosystems, the first priority is to establish a high-level language that is universal, easy to program, and still able to exploit the full performance potential of the hardware," DeepSeek said in the announcement. The company adds that "TileLang was created precisely to meet this need" and that it offers a "simpler programming model" than Nvidia's CUDA. This is DeepSeek's own framing; there is no independent evaluation of TileLang's maturity or characteristics compared with the CUDA ecosystem.
How the pieces fit together
The public TileLang-Ascend repository documents an Ascend-specific NPU intermediate representation (IR) designed to connect the platform to the MLIR compiler ecosystem — an open infrastructure layer for compiler development that is also used across the broader machine learning ecosystem. Huawei's own AscendNPU-IR documentation describes a corresponding TileLang integration for developing high-performance kernels on Ascend processors.
That means TileLang is not merely a language for hand-writing kernels, but an entry point into a standardised compiler layer. For the Ascend platform, it is the path from "hardware that can run models" to "hardware developers can actually build for." The supernode solution based on 128 Ascend 950 chips — which DeepSeek says is optimised for both computation and communication — serves as the hardware anchor for the whole stack.
What remains open
Several significant questions go unanswered in the available evidence:
- Performance. No independent benchmark figures compare the Ascend implementations with Nvidia's versions. Dataquest's article addresses performance measurements, but complete results are lacking. For now, claims that the Ascend implementations "correspond" to the Nvidia components rest on DeepSeek's own wording.
- Maturity. Whether TileLang is in practice "simpler" than CUDA cannot be established from the available sources — that is marketing's choice of words, not an evaluation.
- Source basis. The WeChat post itself is not directly available, and most of the coverage recycles the same Reuters wire. The sources therefore do not independently corroborate one another, and in particular the list of six components rests on a single report.
Context: DeepSeek's broader Huawei bet
The release does not stand alone. According to Bloomberg, cited via CNBC TV18, DeepSeek has previously planned to install at least 160,000 of Huawei's most powerful accelerators in a large data centre the company is building in Inner Mongolia. That is a separate undertaking, not part of this announcement, but it points in the same direction: the company's infrastructure is increasingly being shifted from the Nvidia ecosystem to Huawei.
Taken together, the picture points to a shift in the Chinese AI industry — from improvising ports of models onto domestic chips to building out a parallel software ecosystem underneath them. Whether that software stack measures up in practice remains to be seen.
Sources: Reuters via Yahoo Finance/GuruFocus · Reuters via Deccan Chronicle · Dataquest · Reuters via NDTV Profit · Bloomberg via CNBC TV18