深圳国产芯片成功训练万亿级AI大模型

Odaily Planet Daily News: Recently, the AI training platform project team at Shenzhen He Tao College, in collaboration with Harbin Institute of Technology (Shenzhen), the Shenzhen Institute of Big Data, and relevant teams from Huawei, and alongside the Smart City AI computing platform, has launched a joint effort for large model training on domestic computing power. Leveraging the Ascend 910C domestic AI computing cluster, they successfully completed the full-parameter post-training of the 1.6 trillion parameter model, DeepSeek-V4-Pro.

This practice has accumulated valuable experience for the world's third-party institutions in completing model training of this level on domestic computing power platforms, and also confirms that domestic AI chips can support world-class ultra-large parameter model training. From a technical application perspective, invoking existing models for business inference and completing full-parameter model training from scratch belong to different technical stages, with significant differences in technical difficulty and hardware requirements. The results of this experiment indicate that domestic AI computing power is now capable of undertaking top-tier large model training tasks, and the related technical path is feasible. (Shenzhen Release)