Report says DeepSeek plans to pretrain next-gen V5 on Huawei Ascend chips
CLS reported on November 3 that DeepSeek plans to pretrain its next-generation V5 on Huawei Ascend chips, previously used mainly for inference.
A CLS reporter learned from several industry chain sources that DeepSeek plans to pretrain its next-generation DeepSeek V5 on Huawei Ascend chips, which until now have mainly handled inference for DeepSeek models.
Publicly available information shows the earliest scaled deployment of Ascend for DeepSeek models took place around the 2025 Spring Festival. On January 28, 2025 senior Huawei leadership decided to have Huawei Cloud support DeepSeek rapidly, and between Lunar New Year Eve and the second day of the year Huawei and DeepSeek completed deployment and validation of more than 3,000 Ascend chips, successfully deploying DeepSeek-V3/R1 on the Huawei Cloud Gui'an data center. After DeepSeek-V4 shipped in April 2026, Huawei said that through joint chip-model technology collaboration the full Ascend supernode product line now fully supports the DeepSeek V4 model series.
On September 30, DeepSeek open sourced infrastructure components for the Huawei Ascend compute platform, covering the TileLang compiler toolchain, high performance computing libraries and distributed communication libraries, one to one with the components previously opened for Nvidia platforms. Huawei provided jointly defined Ascend SuperPoD Flex and UBL128 networking solutions and opened the Ascend C programming interface and low level instruction support for TileLang work.
Recently a Shenzhen Hetaizi College AI training platform team, working with Harbin Institute of Technology Shenzhen, the Shenzhen Big Data Research Institute and Huawei teams, completed full-parameter post-training of the 1.6 trillion parameter DeepSeek-V4-Pro on an Ascend 910C domestic cluster, though that training did not involve base pretraining. If DeepSeek next-generation pretraining runs on Ascend compute, it would give a first answer to whether China leading models can complete base pretraining on domestic compute.