DeepSeek open-sources an operator toolchain and ports it to Huawei Ascend

On September 30, 2026, Leiphone reported that DeepSeek had open-sourced a GPU-programming operator toolchain and ported it as-is to Huawei's Ascend 950 NPU, with the stated goal of letting a single codebase run across all chips. According to that report, the toolchain covers the operator development stage and had previously centered on NVIDIA's GPU ecosystem.

A day later, on October 1, InfoQ China described the release in somewhat different terms: DeepSeek had open-sourced infrastructure components for the Ascend platform, spanning TileLang, a compute library and a distributed communication library. That report noted the company had not disclosed further technical details.

On October 2, TMTPost published an analysis of why DeepSeek is embracing Huawei at this point, framing it not as a simple port but as "a move" — with the focus on the timing and the broader shift behind the decision.

The three accounts do not fully agree on what was actually open-sourced: one describes a GPU-programming operator toolchain that runs on the Ascend 950 NPU, while another describes Ascend infrastructure components and names TileLang, a compute library and a distributed communication library. Both mention an open-source release, and neither gives a version number or finer technical metrics. As of October 2, the publicly reported progress remains the open-source release itself, along with discussion of the motives behind it.