DeepSeek V5 plans to pretrain on Ascend chips

Article body

Reporters learned from multiple industry chain sources that DeepSeek's next-generation large model, DeepSeek V5, plans to conduct pre-training based on Ascend chips. Previously, Huawei's Ascend chips were mainly used for inference tasks of DeepSeek large models.

From publicly available information, the earliest large-scale deployment of Ascend carrying DeepSeek models occurred around the Spring Festival of 2025.

Fang Xingdong, Executive Vice Dean of the College of Media and International Culture at Zhejiang University, mentioned in his book "The Rise of Ascend" that on January 28, 2025, Huawei's senior management issued a decision for Huawei Cloud to rapidly support DeepSeek. From New Year's Eve to the second day of the Lunar New Year, Huawei and DeepSeek completed the deployment and verification of more than 3,000 Ascend chips, successfully deploying the DeepSeek-V3/R1 large models to Huawei Cloud's Gui'an Data Center. The book also writes that DeepSeek's explosive growth greatly excited Ren Zhengfei, and throughout the Spring Festival, he was directing Huawei Cloud and related product teams to cope with the traffic tsunami. In January of the same year, SiliconFlow and Huawei Cloud launched DeepSeek-V3 and R1 inference services based on Ascend.

After the release of DeepSeek-V4 in April 2026, Huawei stated that through the two sides' "chip-model technology collaboration," the full series of Ascend supernode products can now fully support the DeepSeek V4 series models. DeepSeek mentioned on its official API page that, limited by high-end computing power, the service throughput of the V4-Pro model is still limited at present, and it is expected that after the Ascend 950 supernodes are mass-produced and launched in the second half of the year, the Pro price will be significantly reduced.

On September 30, DeepSeek announced the official open-sourcing of infrastructure components for the Huawei Ascend computing power platform, covering the TileLang high-level language compilation tool, high-performance computing libraries (DeepGEMM, FlashMLA, TileKernel, DeepSelect, etc.) and the distributed communication library (DeepEP), corresponding one-to-one with the previously open-sourced components for the NVIDIA platform. Huawei provided the jointly defined Ascend supernode SuperPoD Flex and UBL128 networking solutions, and opened the Ascend C programming interface and underlying instruction system to support TileLang's operation.

The significance of this move lies in the fact that DeepSeek is not only using Ascend for inference, but has also systematically aligned and co-built with Ascend at the bottom layer of the software stack, extending adaptation to the infrastructure level that supports training and inference.

Currently, the implementation of domestic computing power is mainly concentrated on the inference side, and the training side still needs improvement. Not long ago, the AI training platform project team of Shenzhen Hetao College, together with relevant teams from Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, and Huawei, in collaboration with the Shenzhen Smart City AI computing power platform, relying on the Ascend 910C domestic AI computing power cluster, completed full-parameter post-training of the 1.6 trillion parameter large model DeepSeek-V4-Pro. It should be noted that this training still does not involve basic pre-training.

The large-scale verification of domestic chips on the inference side provides industry confidence, but the ecosystem migration cost on the training side remains a variable that determines the market ceiling. If DeepSeek's next-generation large model completes pre-training based on Huawei Ascend computing power, it will answer for the first time the question of "whether China's top large models can complete pre-training tasks on domestic computing power," a question that previously had only doubts and no answers.

(Source: Cailian Press) Body middle Embedded advertisement