Shenzhen domestically-produced chips successfully train trillion-level AI large models.

Recently, the Shenzhen Hetao College AI Training Platform project team, in collaboration with Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, and Huawei teams, jointly carried out research and development on domestic computing power for large-scale model training using the Deephi City AI computing platform. With the support of the Ascend 910C domestic AI computing cluster, they completed training on the full parameters of a 1.6 trillion parameter DeepSeek-V4-Pro large model. This practice has accumulated important experience for global third-party organizations to complete training of models at this level on domestic computing platforms, and also confirms that domestic AI chips can support world-class training of super-large parameter models. From a technical application perspective, using existing models for business inference is different from completing full parameter training of models from scratch, with obvious differences in technical difficulty and hardware requirements. The results of this experiment show that domestic AI computing power can now undertake top-level training tasks for large models, and the related technical path is feasible. (Shenzhen releases)