Dotsfeed
← News

DeepSeek open-sources TileLang toolkit for Huawei Ascend chips — China's answer to CUDA

Verified· Sep 30, 2026Published Sep 30, 2026

DeepSeek released six open-source software modules for Huawei's Ascend AI chips, including the TileLang kernel language, on September 30, 2026 — China's latest push for a CUDA alternative.

What happened

DeepSeek announced on September 30, 2026 (via its official WeChat account) that it is open-sourcing programming infrastructure for Huawei's Ascend AI chips, built in close collaboration with Huawei. The release covers six software modules, including computing and communication libraries, building on infrastructure DeepSeek previously developed for Nvidia hardware. The key component is TileLang, an open-source programming language for high-performance AI kernels that DeepSeek describes as a simpler programming model than Nvidia's CUDA; an Ascend-compatible version now officially supports Huawei's Ascend 950 accelerators. Also released: DeepGEMM Ascend (matrix-multiplication kernels) and DeepEP Ascend (large-scale cross-device communication), plus vector, sparse-attention, and data-filtering modules. DeepSeek says every TileLang operator used in its V4-series training now has a high-performance Ascend implementation, and the same Python interfaces work across Nvidia GPUs and Huawei NPUs with automatic backend selection. The two companies also jointly optimized a 128-chip Ascend 950 supernode for large-scale workloads.

Why it matters

This is China's most concrete step yet toward software independence from Nvidia's CUDA ecosystem, arriving as US export restrictions keep cutting off China's access to advanced Nvidia processors. Chip production remains a separate constraint, but DeepSeek just materially narrowed the software gap on Huawei hardware — and it's all open source and free to download.

Sources

Verified September 30, 2026.

Sources

DeepSeek open-sources TileLang toolkit for Huawei Ascend chips — China's answer to CUDA · Dotsfeed