DeepSeek open-sources TileLang AI chip tools for Huawei's Ascend accelerators

Editorial illustration: Translucent turquoise puzzle pieces extend from an open metal case into a recessed opening in a large dark processor on a circuit board.

In brief

  • DeepSeek announced the open-source toolkit on September 30 via its WeChat account.
  • TileLang leads the release, alongside five libraries including DeepGEMM and FlashMLA.
  • The software is reportedly optimized for Huawei's Ascend 950 chips and 128-chip supernodes.
  • Crypto Briefing framed the release against China's reliance on Nvidia's CUDA.

What's in the release

TileLang isn't shipping alone. DeepSeek packaged a collection of compute and communication libraries alongside it (DeepGEMM, FlashMLA, TileKernel, DeepSelect and DeepEP), according to Crypto Briefing's account of the WeChat announcement. Each library covers a different layer of AI development, from matrix multiplication to memory management to inter-chip communication, the report said.

There's benchmarking in the bundle, too. The release includes tools meant to help developers with kernel development on the Ascend architecture.

On hardware, the announcement as reported says the software is optimized for Huawei's Ascend 950 chips and supports a "supernode" configuration that links 128 of those accelerators. Huawei reportedly provided extensive support during development, per Crypto Briefing.

Why Crypto Briefing ties it to CUDA

The outlet's framing leans heavily on Nvidia. It said the vast majority of AI models, training pipelines and inference systems worldwide are built on CUDA, and it described that dependency as a strategic vulnerability for China after the US imposed export controls on Nvidia's most advanced chips.

That's the backdrop Crypto Briefing uses for TileLang.

It's the publication's analysis, and it's worth reading that way. The WeChat announcement, as reported, describes a set of tools, libraries and benchmarks for one chip family; the broader argument about China's reliance on Nvidia comes from the outlet.

A continuing Huawei track

This isn't DeepSeek's first move toward Huawei silicon. In April 2026, the company adapted its V4 AI model to run on Huawei hardware, Crypto Briefing reported. The new toolkit extends that work from a single model to a set of tools developers can download for free.

Crypto Briefing also reported, without citing a source, that DeepSeek plans a large data center in Inner Mongolia designed to host at least 160,000 Huawei Ascend accelerators. DeepSeek's announcement, as described in the report, doesn't include those figures.

Frequently asked questions

What is TileLang?

TileLang is a high-level programming language included in DeepSeek's open-source toolkit for Huawei's Ascend AI accelerators. Crypto Briefing described it as a domestic counterpart to Nvidia's CUDA. The tools are available for free download.

Which Huawei chips does DeepSeek's new software support?

According to the announcement as reported by Crypto Briefing, the software is optimized for Huawei's Ascend 950 chips. It also supports a supernode configuration linking 128 of those accelerators.