DeepSeek-V4 Model Shows Performance Gains on Huawei AI Infrastructure

A new research paper details an optimization framework that significantly boosts the performance of trillion-parameter AI models on Huawei's Ascend NPU SuperPOD. This framework, applied to DeepSeek-V4, achieved a substantial improvement in Model FLOPs Utilization and surpassed other models in operations-research tasks. The development highlights the advancement of a fully Chinese AI training stack independent of Nvidia technology.