China's own AI chips now power massive 1.6-trillion-parameter model training

A team from Shenzhen's He Tao College AI training platform, together with Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, and Huawei, has used Huawei's Ascend 910C chips to fully train the DeepSeek-V4-Pro model, which has 1.6 trillion parameters. This is the first time a third party has completed such a large-scale model training on China's own AI computing platform. The project achieved over 30% model flops utilization (MFU) and a 14% boost in key training operator efficiency, meeting industrial standards. Experts say this proves China's AI chips can now handle top-tier model training tasks.