Huawei's Ascend 910C Completes Post-training of DeepSeek's 1.6 Trillion-Parameter Model with 1,000+ GPUs

According to Beating, a joint team from Shenzhen Hetao College, Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, Huawei, and Deepcity AI has announced the successful completion of full-parameter post-training for DeepSeek-V4-Pro, a 1.6 trillion-parameter model, on domestic AI infrastructure. This marks the first time a third-party organization has completed full-parameter post-training for a model of this scale on Chinese hardware.

The team leveraged a cluster of over 1,000 Huawei Ascend 910C chips to overcome communication bottlenecks through optimized distributed load balancing. During the 1,500-step training process, the system operated without interruption, achieving a model FLOPs utilization (MFU) rate exceeding 30% and improving key operator efficiency by 14%, meeting industrial-grade performance standards.