Event date · · DeepSeek

Reportedly: DeepSeek Open-Sources Ascend Versions of TileLang, DeepGEMM and DeepEP as Huawei Details SuperPoD Flex

DeepSeek 深度求索Chinese AIOpen weights
FACT STATEMENT

DeepSeek released Ascend versions of TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect, while Huawei detailed a jointly defined 128-card SuperPoD Flex design. The source announcement was published on 2026-09-30, Pandaily reported.

China context

Original name
深度求索
Outside China
Open weights · pandaily.com
Claims
Company-reported; not yet independently evaluated
For builders
Developers can now access Ascend-compatible implementations of DeepSeek's core libraries, enabling experimentation with Huawei hardware without starting from scratch.
For investors
The collaboration between DeepSeek and Huawei on SuperPoD Flex signals a maturing domestic AI infrastructure stack, which may affect the competitive landscape for AI hardware and software providers.
What happened

DeepSeek has open-sourced Ascend-adapted versions of several core infrastructure components: TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Huawei also provided details on a jointly defined 128-card SuperPoD Flex design. The release was reported by Pandaily on 2026-09-30.

Technical significance

The open-sourcing of Ascend versions of TileLang, DeepGEMM and DeepEP indicates that DeepSeek's training and inference stack can now run on Huawei's Ascend hardware. The 128-card SuperPoD Flex design suggests a scale-out architecture for large model training. Verification of actual performance parity with NVIDIA-based systems would require independent benchmarks.

Industry impact

Developers outside China who rely on NVIDIA CUDA now have a reference implementation for porting DeepSeek's stack to Huawei Ascend, potentially reducing lock-in and enabling deployment on alternative hardware. Huawei gains a validated software stack for its Ascend chips, strengthening its position in the domestic AI infrastructure market.

Decision value

For enterprises considering Huawei Ascend hardware, this release lowers the software barrier to running DeepSeek models. It may also influence procurement decisions in markets where NVIDIA GPUs are restricted or expensive.

What to watch

A concrete next signal is whether DeepSeek publishes performance benchmarks comparing Ascend and NVIDIA implementations of these components. Another is whether third-party developers adopt the Ascend versions for production workloads.

Latest in Chinese AI

  1. MetaXReportedly: MetaX Adds Hy4 Support to C-Series GPUs With FlagOS W8A8 Weights
  2. XiaomiReportedly: Xiaomi adds MiMo UltraSpeed to Token Plan Pro and Max for a limited time
  3. MiniMaxMiniMax publishes cryptanalysis findings from its internal AI model
  4. MiniMaxMiniMax publishes Cyber repository on GitHub
  5. TencentReportedly: Tencent opens Hy4 weights, completing the Hunyuan LLM open-weight rollout

All China AI Events

AIGC Newsletter

China AI, with sources and context.

Analysis of Chinese AI models, companies and policy, and what you can use outside China.