DeepSeek ports its core kernel libraries to Huawei's Ascend 950
DeepSeek published Huawei Ascend 950 versions of DeepGEMM, its matrix multiplication library, and DeepEP, its expert-parallel communication library, on GitHub. Both keep the same APIs as their NVIDIA versions.
DeepSeek opened two new repositories on GitHub on September 30, 2026: DeepGEMM-Ascend and DeepEP-Ascend. The company has ported the core kernel libraries it uses to train and run its models to Huawei's Ascend 950 AI processors. There was no official blog announcement; the details come from the repositories' README pages.
What the two libraries do
DeepGEMM is a kernel library that speeds up matrix multiplication (GEMM), the basic operation of large language models. According to the README, DeepGEMM-Ascend is fully API-compatible with DeepGEMM: developers can install the package and use the same calls and workflow on Ascend. The library supports matrix multiplication in BF16, FP8 and FP4, MegaMoE kernels for mixture-of-experts (MoE) models, and kernels for MQA attention. The repository is released under the MIT license.
DeepEP is the communication library used in MoE models to dispatch data to experts on different processors and combine the results. DeepEP-Ascend's public APIs are aligned with the NVIDIA version. The library also includes FP8 dispatch, send and receive operations for pipeline parallelism, and an experimental remote memory access component.
Measurements and driver timeline
The performance results DeepSeek shared were measured on the Ascend 950DT. According to the DeepEP-Ascend README, they were taken with a proof-of-concept driver package (PoC HDK) that Huawei supplied to DeepSeek, plus additional manual configuration, and that package is not publicly distributed. The commercial driver release with the configuration needed for full bandwidth is planned for around October 15 on Huawei's Atlas 850E software page. DeepSeek notes that the date is Huawei's plan and may change.
A major step for non-CUDA hardware
Most AI software is written for NVIDIA's CUDA ecosystem, and moving to other hardware often gets hard at that software layer. DeepSeek, whose open models are widely used, porting its own kernel libraries to Ascend with the same APIs means significant software support for hardware outside CUDA. The move also fits China's effort to build its own AI hardware chain under US export restrictions. Teams that want to train or run MoE models on Ascend 950 systems will be able to move code written for NVIDIA with fewer changes.
The performance results are DeepSeek's own measurements taken with a driver configuration that has not been released yet, and results with the commercial driver may differ. Support on other Ascend generations and CANN versions has not been established either. The repositories are brand new, and it is worth following evaluations from independent researchers.
Sources: GitHub: deepseek-ai/DeepGEMM-Ascend, GitHub: deepseek-ai/DeepEP-Ascend.
