Recently, the integration of domestic large models and domestic computing power has taken another step forward. According to sources within the industry, DeepSeek's next-generation large model, DeepSeek V5, plans to use Huawei's Ascend 950 chip as its core training power, with the project expected to be implemented next year. So far, DeepSeek has not publicly responded to this matter. According to insiders, the pace of collaboration between the two parties has exceeded external expectations, with the core R&D teams of DeepSeek and Ascend already stationed in Hangzhou, having completed a significant amount of underlying operator adaptation and custom development during the V4 phase.
A more critical change lies in the division of labor: in the previous version of DeepSeek, Ascend's computing power was only responsible for inference tasks and did not participate in the training phase; however, V5 will be the first large model to natively rely on the Ascend 950 for training. In an industry environment where the supply of high-end AI chips remains tight, leading domestic large model manufacturers are accelerating their transition to domestic computing power bases. However, the difficulty of truly transitioning from inference to training is not on the same level: training requires higher standards for operator completeness, cluster interconnectivity, and long-term stability, and any shortcoming could amplify uncertainties in the entire training task.
DeepSeek V5 will be the first large model to natively rely on the Ascend 950 for training
Looking back at DeepSeek's recent product line, it is evident why stronger computing power support is necessary. On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017. In the DeepSWE v1.1 test, which measures software engineering capabilities, this model achieved a Qualcomm pass rate of 77.0%, outperforming mainstream overseas models such as GPT-6 Astra and Claude Opus 5.H
About the author
Henderson

