- The open-source toolkit spans compiler tools, computing kernels and distributed communications for Huawei’s Ascend platform
- The move extends DeepSeek’s Ascend support from model deployment to training, addressing a key weakness in China’s AI-chip ecosystem
DeepSeek on September 30 open-sourced a suite of infrastructure components for Huawei’s Ascend AI chips, extending its support for the domestic platform from inference deployment to model development and training.
The release includes TileLang, a high-level programming and compilation tool, along with computing and distributed-communication libraries.
The components correspond to DeepSeek’s previously open-sourced tools for Nvidia’s platforms, giving developers an equivalent software stack for Ascend.
TileLang is particularly important because it allows developers to write high-performance computing operations at a higher level of abstraction than Nvidia’s CUDA programming language.
DeepSeek expands its stack
Instead of writing low-level instructions directly, developers can use TileLang to describe how computations should be organized while the compiler translates them into code that can run efficiently on the underlying hardware.
DeepSeek said the operators it uses to train its V4 models with TileLang now have corresponding high-performance implementations on Ascend chips.
In effect, the company has ported a core part of its model-development toolchain from Nvidia’s ecosystem to Huawei’s.
From deployment to training
The release also includes DeepGEMM for accelerated matrix multiplication, DeepEP for large-scale communication between devices, TileKernels for common vector and memory operations, FlashMLA for more efficient long-context processing and DeepSelect for data selection.
DeepSeek said several of the components have reached performance close to hardware limits in testing.
Huawei and DeepSeek have also worked on a 128-card supernode based on Huawei’s Ascend 950, with joint optimization of computing and communications, according to DeepSeek.
The company credited Huawei’s team with providing what it described as strong and extensive support during development.
Closing the software gap
The shift comes as China’s AI-chip ecosystem seeks to close a software gap with Nvidia.
Until now, much of DeepSeek’s Ascend work had focused on deploying models for inference. Supporting the tools used to develop and train models goes a step further by making it easier to build future models on domestic hardware.
Huawei has also been opening up its own software stack. At its 2026 Huawei Connect conference in September, the company announced broader open-source access to CANN, or Compute Architecture for Neural Networks. This is the software layer that connects Ascend chips with AI developers and applications.
Breaking through the chip barriers
The push comes as access to high-end Nvidia chips in China remains constrained by US export controls. At the same time, Huawei is positioning newer Ascend chips as alternatives for domestic AI workloads.
DeepSeek’s V4 model has already listed both Ascend and Nvidia hardware in its validation work, a move that underscored the growing importance of Huawei’s platform.
Nvidia CEO Jensen Huang has warned that a future in which DeepSeek develops its models first on Huawei hardware would be a serious setback for the US.
“The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation,” he said while speaking on the Dwarkesh Podcast in April 2026.
DeepSeek’s latest open-source release addresses one of the most difficult parts of that transition: not just making Chinese AI chips usable, but making them easier for developers to build on.
