Rumor has it that DeepSeek V5 is expected to complete training based on the Huawei Shengteng 950 chip next year

The Zhitong Finance App learned that, according to reports, some media learned from people familiar with the industry chain that DeepSeek V5, the next big model of DeepSeek, plans to use the Huawei Ascend 950 chip as the core training computing power, and the pace of the project points to implementation in 2027.

According to reports, at present, DeepSeek's core R&D and matchmaking team with Huawei Shengteng is based in Hangzhou, and the two sides have completed a large number of low-level operator adaptations and customized development in the V4 phase, laying the foundation for subsequent model training migration and computing power collaboration.

Judging from the cooperation path, in the previous DeepSeek model, Huawei's Shengteng computing power mainly undertook inference tasks and has not yet entered the model training process. However, the current V5 plan introduces the Shengteng 950 into the core training system, which means that the cooperation between the two sides is expected to extend further from the reasoning side to the training side. Compared to inference tasks, large model training places higher demands on chip computing power, communication interconnection, software ecology, and cluster stability.

The industry believes that if DeepSeek V5 finally achieves native training based on the Shengteng 950, it is expected not only to further verify the performance and adaptability of domestic AI chips in large-scale model training scenarios, but also bring new commercialization opportunities to the domestic computing power industry chain. However, the relevant cooperation is currently still based on media citing information from people in the industry chain. The specific technical plan and implementation progress are yet to be further confirmed by the government.