DeepEP-Ascend and DeepGEMM-Ascend
DeepSeek ports its MoE communication and FP8 GEMM libraries to Huawei Ascend NPUs, targeting Ascend 950 with CANN 9.2.0.
First-party Ascend-native versions of DeepEP and DeepGEMM; DeepEP-Ascend reports dispatch at roughly 90-95% of physical bandwidth for EP size up to 32 on Ascend 950DT. Shows DeepSeek serving on domestic Chinese silicon.
- Date
- Tuesday 29 September 2026
- Lab
- DeepSeek
- Kind
- infra
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| EP dispatch efficiency | 90-95% of payload bandwidth limit EP size up to 32, Ascend 950DT, CANN 9.2.0 | company |
Dates are GitHub repo creation (DeepEP-Ascend 2026-09-30, DeepGEMM-Ascend 2026-09-29). Repo says full-bandwidth Atlas 850E HDK is planned for mid-October 2026. Not a model release; no statement on whether V4 trains on Ascend.
Sources
- github.com/deepseek-ai/DeepEP-Ascend
- api.github.com/orgs/deepseek-ai/repos?sort=created&direction=desc&per_page=30
- github.com/deepseek-ai/DeepGEMM-Ascend
- gigazine.net/gsc_news/en/20261001-deepseek-huawei-ascend/
- forklog.com/en/deepseek-unveils-open-source-software-ecosystem-for-huawei-ai-chips/
- tomshardware.com/tech-industry/artificial-intelligence/deepseek-and-huawei-release-open-so
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check