Breaking

DeepSeek Expands Huawei Ascend Stack With Six Open-Source AI Infrastructure Projects

September 30, 2026 at 14:47 EDT

Infra & Chips
Open Source
Software Dev & Coding

DeepSeek has published or updated a six-part open-source infrastructure stack for Huawei Ascend AI accelerators, moving more of its model-serving and kernel work toward a non-Nvidia software path. The release spans TileLang, DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA and DeepSelect, with Huawei support acknowledged across the Ascend-specific engineering work.

Sep. 30, 2026 · DeepSeek

DeepSeek opens an Ascend AI software stack

Six open-source infrastructure projects now target Huawei Ascend chips, from a kernel language to matrix math, communication and sparse-attention operators.

projects in the Ascend stack

99.8%

reported DeepGEMM-Ascend hardware utilization

375

GB/s upper EP8 dispatch bandwidth

2-20x

DeepSelect speedup in supported TopK cases

Reported Ascend performance versus hardware peak

99.8%

DeepGEMM

95%

FlashMLA prefill

83%

FlashMLA decode

The strongest numbers are near-peak matrix math and sparse-attention prefill; decode and larger communication modes are still being optimized.

Stack map

TileLang→DeepGEMM-Ascend→DeepEP-Ascend→FlashMLADeepSelectTileKernels

Why builders care

Aligned APIs and automatic backend selection can reduce porting work for DeepSeek-style kernels on Ascend systems.

What still limits rollout

Some paths remain CUDA-only, earlier model support changed, and the public firmware baseline is still pending.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.