On June 5, reports emerged indicating that the AI training platform project team from the Shenzhen Futian Higher Education Cluster recently collaborated with Harbin Institute of Technology (Shenzhen), the Shenzhen Research Institute of Big Data, and pertinent teams from Huawei, as well as leveraging the AI computing power platform of the Shenzhen Smart City Development Group. Together, they took on the challenge of large-scale model training using computing power produced in China. Utilizing the domestic AI computing cluster Ascend 910C, the project achieved the full-parameter post-training of the massive 1.6-trillion-parameter model, DeepSeek-V4-Pro. This accomplishment not only offers invaluable insights for global third-party organizations looking to train similarly sized models on computing platforms produced in China but also underscores the capability of domestic AI chips to support the training of world-class, ultra-large-parameter models.

From a technical standpoint, deploying existing models for business inference and training models from scratch with full parameters are two distinct technical stages. They differ significantly in terms of technical complexity and hardware demands. The outcomes of this experiment reveal that AI computing power produced in China is now competent to handle top-tier large model training tasks, and the relevant technical approach is viable.