DeepSeek has open-sourced six software modules built for Huawei's Ascend AI chips, releasing an Ascend-compatible version of its TileLang programming language alongside kernel and communication libraries that the Chinese start-up says approach the hardware's performance limits in several tests.

The Hangzhou-based company announced the releases Wednesday on its official WeChat account. The modules mirror DeepSeek's existing open-source tools for Nvidia GPUs and aim to build what the company called an "independent and controllable" software ecosystem for Chinese accelerators, a direct response to Washington's export controls that block Nvidia's most advanced AI chips from the Chinese market.

The centerpiece is TileLang, a custom programming language designed to streamline development of high-performance kernels — the essential computation programs that run on GPUs and CPUs. TileLang still lists Nvidia as its primary back end, but a project update on GitHub confirms official support for Huawei's Ascend 950 accelerators, including "native code generation, automatic scheduling, and synchronisation."

"As an open-source project, the TileLang Ascend version aims to serve as an example for building a highly available software ecosystem for more AI chips," DeepSeek said.

Beyond TileLang, DeepSeek published a set of underlying libraries tailored for high-throughput training and inference on Ascend silicon. DeepGEMM Ascend is a matrix multiplication kernel library, one of the most compute-intensive operations in neural network workloads. DeepEP Ascend handles efficient large-scale cross-device communication, a critical bottleneck when training models across thousands of chips. Additional modules target long-context processing efficiency and data filtering on Huawei hardware.

DeepSeek said the tools delivered computational and communication performance approaching "the hardware limits" in several tests. The company did not publish detailed benchmark figures.

The infrastructure work stems from what DeepSeek described as "close collaboration" with Huawei. The two companies jointly optimized a supernode architecture powered by 128 Ascend 950 processors — a configuration that gives Chinese model developers an alternative to the large-scale GPU clusters that Nvidia historically dominated.

The partnership has deepened over the past year. When DeepSeek launched its V4 model in April, Huawei quickly announced "full support" for inference workloads — the process of running a trained AI model to make predictions on new data — using its own chips. Other Chinese chip designers have followed: Cambricon Technologies, Moore Threads and Alibaba Group Holding's T-Head have all confirmed compatibility with the DeepSeek model. Alibaba owns the South China Morning Post, which first reported the releases.

Huawei, for its part, is accelerating its silicon roadmap. Earlier this month the company said the Ascend 960DT training chip would be "ready in the first quarter of 2027," pulling the schedule forward by three quarters from its original target.

The software gap has long been Huawei's biggest handicap against Nvidia, whose CUDA ecosystem locks in developers across the industry. DeepSeek's open-source releases attack that moat directly, giving Ascend hardware a kernel development stack that outside engineers can inspect, extend and port to other domestic chips.

If Huawei delivers the 960DT on its new schedule and the Ascend software stack matures at the current pace, Nvidia's remaining share of the Chinese AI accelerator market will face a genuinely domestic challenger for the first time since export controls began.