DeepSeek releases Ascend tools as developers weigh alternatives to Nvidia

DeepSeek has published two new open-source repositories for Huawei’s Ascend accelerators: DeepGEMM-Ascend for matrix multiplication and DeepEP-Ascend for the all-to-all communication used by mixture-of-experts models. Its announcement also included Ascend updates to existing projects TileKernels, DeepSelect and FlashMLA.

TileLang, described by Bloomberg as a centerpiece of the effort, is neither new nor a DeepSeek project. It was open-sourced in January 2025. Ascend adapters were published in the separate tilelang-ascend repository on September 29, 2025, before the recent announcement. TileLang supports both Nvidia’s CUDA and Huawei’s Ascend as backends.

The new repositories have specific hardware and software requirements. DeepGEMM-Ascend requires Ascend 950-series chips and CANN 9.20. DeepEP-Ascend requires Ascend 950 silicon with UBMEM connectivity, limiting its potential audience to those with access to that hardware.

The projects aim to reduce the effort involved in switching: DeepGEMM-Ascend retains the Python package name and API shape of its CUDA counterpart, while DeepEP-Ascend aligns its public buffer interfaces with DeepEP. GitHub figures cited in the source show 515 stars and 39 forks for DeepGEMM-Ascend, compared with 8,522 stars and 1,359 forks for DeepGEMM. DeepEP had more than 10,200 stars, versus 224 for DeepEP-Ascend.

Reporting cited by TechRadar describes past usability concerns with Ascend and CANN. The source also says Huawei has sent engineering teams to customers including Baidu and Tencent to help port CUDA training code. The new repositories provide production kernels, but the source notes they run on one chip family and toolkit version, with firmware that may be difficult to access outside China.