DeepSeek on Huawei Ascend: Is TileLang a CUDA Alternative?

DeepSeek’s Ascend Tools: Six Utensils for Another Stove

Think of an AI model as a recipe and a chip as a stove. For more than a decade, most recipes have been written for one brand of stove, using that brand’s own utensils. Utensils here means everything between the recipe and the flame, including the notation the recipe is written in. On 30 September, DeepSeek published a set for a different stove: open-source software for Huawei’s Ascend chips, built around a language called TileLang. Some coverage framed DeepSeek’s Ascend tools as a step toward a CUDA alternative.

DeepSeek released tooling, the code that sits between a model and a chip. The model is unchanged and the chip is still Huawei’s own design. Six weeks on, the repositories tell a narrower story than the headlines did.

Nvidia, CUDA, Ascend and the Software Stack in Plain Words

If you’ve written a little code but never touched a chip, here is the minimum.

The stack is why chips are hard to swap. You can buy an Ascend board. If your code only speaks CUDA, the board can’t run it, and a chip with a layer missing is a circuit board with its lights on.

What Did DeepSeek Open-Source for Huawei Ascend?

DeepSeek announced the release on its WeChat account. Reuters reported it the same day as six software modules that mirror tools DeepSeek had already published for Nvidia chips, a report summarised by Tom’s Hardware. DeepSeek has not published a numbered list of the six in English, so the table below is our reading of the coverage and the public repositories.

Three of those names are old. FlashMLA, DeepEP and DeepGEMM first appeared in February 2025 for Nvidia chips, and the DeepGEMM-Ascend repository describes itself as a port that is fully compatible with the original. So this is DeepSeek’s existing toolbox, rebuilt for Ascend.

Huawei’s role is narrower than some headlines suggest. DeepSeek said Huawei “fully supported” the work, and the two companies tuned a cluster of 128 Ascend 950 chips together, according to The Decoder. The repositories thank Huawei for “technical support and engineering expertise”. They are published under DeepSeek’s name, with Huawei’s engineers helping.

What Is TileLang?

TileLang is a language for writing kernels. It looks like Python and sits on top of Apache TVM, an open-source compiler, which turns a description of a calculation into code for a particular chip. Its project page lists backends, meaning compilers for a particular chip, for Nvidia, AMD and, since 30 September, the Ascend 950.

It isn’t DeepSeek’s language. The project credits its first version to three developers supervised by Prof. Zhi Yang at Peking University, and it has been public since January 2025. DeepSeek had been using it for about a year before this release, The Decoder reports. Adapters for older Ascend chips have sat in a side project since September 2025. What arrived on 30 September this year was official support for the 950, with DeepSeek’s libraries on top.

TileLang vs CUDA

Several projects, ZLUDA among them, try to make existing CUDA code run on other chips. TileLang goes around CUDA: the kernel is written in a language no chip maker owns, and the project supplies a compiler for each chip. DeepSeek’s TileKernels notes say “the same Python APIs run on both NVIDIA GPUs and Huawei NPUs” (NPU is Huawei’s name for its AI chips). The notes don’t say how much of the code underneath is shared. If the approach holds up outside DeepSeek, the recipe notation stops belonging to one stove maker.

Why Did DeepSeek and Huawei Both Need This?

AI software grew up on CUDA; Nvidia marked the platform’s twentieth year in March. Huawei had chips and its own stack, and it said in August 2025 that it would open CANN’s source, a move we looked at in our piece on the Ascend 960. What it had not had, as far as we can tell, was a leading AI lab publishing how it runs real workloads on Ascend. That takes people who train models and know which five calculations eat the time.

For DeepSeek, the likeliest motive is supply. Chinese labs’ access to Nvidia hardware is squeezed from two sides. Analyst Patrick Moorhead told CNBC on 10 November that Washington has licensed sales of Nvidia’s H200 chips to Alibaba, Tencent, ByteDance and others, but that Beijing hasn’t approved large-scale imports. A lab that can’t count on buying Nvidia chips needs its software to run on something else. A software stack makes the chips a lab can get usable; whether it makes them competitive is unproven.

Then there’s the choice to publish. A toolset only one company uses attracts few outside programmers, and its bugs get found by one team. Open-sourcing works as recruitment: it invites outside programmers to try Ascend, and to find the bugs. It fits the pattern we described in China’s open-source AI strategy.

Are DeepSeek’s Ascend Tools a CUDA Alternative Yet?

Not yet, and the repositories say why more plainly than the coverage did.

The speed figures measure the chip against itself

DeepSeek reports that DeepGEMM-Ascend reaches up to 99.8% of the hardware limit, the most arithmetic the chip can do in theory, on an Ascend 950DT, one model in the 950 family. Its FlashMLA kernels reach 95% of the chip’s peak when reading a prompt and 83% when generating. These figures measure how little of the chip the software wastes. A comparison with Nvidia would need a different test, and we found no independent benchmark of this stack.

It needs Huawei’s newest chip

The Ascend side of every component targets one chip generation, the 950. The DeepEP-Ascend notes go further: the tests ran on a hand-configured proof-of-concept setup, and support on other Ascend generations “has not been established”. So running this today means having an Ascend 950, and Huawei has not published how many are in customers’ hands.

One of the six has no licence file

Five of the six carry the MIT licence, which lets anyone reuse the code, including in commercial products. DeepEP-Ascend has no licence file at all, and a request to add one has been open since the day of release. Code without a licence can generally be read but not legally reused. This is also the piece that moves data between chips, which is what you need for anything larger than a single machine, and its notes say parts of it “remain under optimization”. The missing file looks like an oversight, but six weeks is a long oversight.

Existing CUDA code doesn’t move itself

Open tools make the first step cheaper, and the migration is still each team’s own work. Someone has to rewrite the heavy calculations, check the results against the CUDA version and measure the speed on the new chip. For a company with years of CUDA code, that cost is the same as it was in August.

CUDA’s real advantage is people

Nvidia’s own materials this year put CUDA’s developer base at “over 4 million” in one document and 6 million in another. Huawei’s figure for CANN at its September conference was a little over 5,200 monthly active developers. Universities, textbooks and job adverts all speak CUDA. Six repositories don’t change that; years of people using them might.

Nvidia also has a counter-argument. Asked at its GTC conference whether AI that writes code would weaken CUDA, Ian Buck, the Nvidia vice-president who runs its data-centre computing business, said agents are already writing and tuning CUDA kernels. “It’s actually accelerating CUDA adoption,” he said, as Tech Wire Asia recounted in October. The same tools can be pointed at TileLang too, so that argument may cut both ways.

What Has Changed Since September?

Both DeepEP-Ascend (3 November) and FlashMLA (27 October) have been updated since launch, though the licence request is still unanswered.

A bigger claim arrived on 3 November: China’s STAR Market Daily said DeepSeek plans to train its next model, V5, from scratch on Ascend chips, according to a summary by the newsletter Geopolitechs; we have not seen the original, and neither DeepSeek nor Huawei has confirmed it. Running a finished model for users is called inference. Training one from scratch is far more demanding, and these six tools are the groundwork such an attempt would need.

What Does This Mean If You Buy or Build AI?

When a headline says an AI company “open-sourced” something, ask which of three things it was: the model, the tools or the chip. This time it was tools. Then:

• Choosing hardware? There is now one published example of an AI lab running its own stack on Ascend. It was done on the newest chip, with Huawei’s engineers helping. That shows it can be done; it’s no measure yet of what it costs anyone else.
• If you write model code, three questions decide whether a stack is ready for production. Does it work with the framework you already use, such as PyTorch? Do the results match your CUDA version? How does the speed compare, and on which calculations? Test your single heaviest calculation first. If the results don’t match, it’s too early. If they match, speed and price are what’s left to argue about.
• If you use AI through an app or an API, nothing changes this quarter. Over a longer stretch, a second usable source of computing is what tends to move prices, which is why chip stories like Tencent’s attempt to rent Nvidia chips through Oracle are worth following even if you never buy a server.

If you’re working out which AI tools and suppliers your own business should depend on, that’s part of our AI agent development work.

Will Programmers Change Stoves?

Nvidia’s position is unchanged for now. DeepSeek showed its own models can run on a second stove and gave away the utensils. One repository still has no licence, and the chip is Huawei’s newest. Watch for an outside team publishing results on Ascend with these tools, or for the V5 report being confirmed. Free tools are the easy part. Will programmers move?

FAQ: DeepSeek, Huawei Ascend and TileLang

What did DeepSeek open-source for Huawei Ascend chips?

On 30 September 2026 DeepSeek published open-source software that lets its own AI workloads run on Huawei’s Ascend 950 chips. Coverage citing Reuters counted six modules. The public repositories show an Ascend 950 compiler for the TileLang language plus TileKernels, DeepGEMM-Ascend, DeepEP-Ascend, FlashMLA and DeepSelect. It is tooling only; the models and the chip design are unchanged.

What is TileLang?

TileLang is an open-source programming language for writing kernels, the small, heavily tuned routines that do the arithmetic inside an AI model. It has Python-like syntax and is built on the Apache TVM compiler. It was first developed by a group supervised by Prof. Zhi Yang at Peking University and has been public since January 2025. It has compilers for Nvidia, AMD and, since 30 September 2026, Huawei’s Ascend 950.

TileLang vs CUDA: what is the difference?

CUDA is Nvidia’s own software and programming model, and code written for it runs on Nvidia hardware. TileLang is an open-source, Python-like language with compilers for Nvidia, AMD and Huawei Ascend 950 chips. Existing CUDA code still has to be rewritten by hand to use it. CUDA has several million developers, by Nvidia’s count; TileLang’s user base is still small.

Are DeepSeek’s Ascend tools a CUDA alternative?

For DeepSeek’s own models on Huawei’s newest chip, they work as one. For most other teams they are an early option: they need an Ascend 950, one of the six repositories has no licence file, and existing CUDA code has to be rewritten and re-tested. CUDA also keeps its libraries, tutorials and several million developers.

Did DeepSeek and Huawei build the tools together?

The repositories are DeepSeek’s. Their notes thank Huawei for technical support and engineering expertise, and DeepSeek said Huawei fully supported the work, according to The Decoder, citing Reuters. The two companies also tuned a cluster of 128 Ascend 950 chips together, The Decoder reports. Neither company has announced a joint product or a formal agreement that we could find.

Are DeepSeek’s Ascend tools faster than CUDA on Nvidia chips?

We found no published comparison. DeepSeek’s figures, such as matrix multiplication reaching up to 99.8% of the hardware limit on an Ascend 950DT, measure how much of that chip’s own capacity the software uses. A comparison with an Nvidia chip would need a different test, and we found no independent benchmark as of 11 November 2026.