DeepSeek plans to train V5 on Huawei Ascend chips, STAR Market Daily reports

DeepSeek plans to train its next-generation model, DeepSeek V5, on Huawei Ascend chips, China's STAR Market Daily reported on November 3, citing several people in the industry supply chain. DeepSeek has not confirmed it. We read the report as republished on The Paper.

WHAT THE REPORT SAYS

- V5 is planned to train on Ascend. Until now, per the report, Ascend chips mainly ran inference for DeepSeek's models. - It points to groundwork already laid: at the end of September DeepSeek open-sourced infrastructure components for Huawei's Ascend platform, covering the TileLang compiler, compute libraries and distributed communication libraries, each matching a component it had already released for NVIDIA hardware. - The report is a short one, by STAR Market Daily reporter Huang Xinyi, and names none of its sources.

WHAT IS SOURCED ELSEWHERE

- October 20: DeepSeek open-sourced DeepInfer, the inference engine behind V4.1 Flash, with two hardware targets: NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs. We read that repository directly. - October 19: The Information reported, citing two people familiar with the matter, that DeepSeek is working closely with Huawei to shape the design of Huawei's future AI chips.

WHY IT MATTERS

Inference on domestic chips is common now. Pre-training a frontier model on them is the hard part, and it is the step that would show Chinese accelerators can carry a lab's whole pipeline rather than only serve the finished model.

WHAT WE DO NOT KNOW

- Whether V5 pre-trains entirely on Ascend, or splits the work with NVIDIA hardware. - Which Ascend generation, how many chips, and when training starts. - When V5 ships, and its size, price or license. DeepSeek has announced no V5 and no date; its next named release is still V4.1 Pro.

Posts on X first flagged the report; the details above are from STAR Market Daily's own text. ArtificialWatch's DeepSeek V5 hunt stays armed and alerts the minute a V5 id appears on DeepSeek's API or the catalogs we sweep.