According to Dongcha Beating monitoring, a joint research team composed of Shenzhen Graduate School, Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, and a Huawei-related team, in collaboration with the Shenzhen AI Computing Power Platform, announced the successful full-parameter post-training of the 1.6 trillion-parameter large model DeepSeek-V4-Pro on a domestic AI computing power platform.
This marks the first time a third-party organization globally has completed full-parameter post-training of a 1.6 trillion-parameter model on a domestic computing power platform.
Compared to pre-training from scratch, the post-training stage (mainly including supervised fine-tuning SFT and reinforcement learning RL) focuses on aligning the model with high-quality instructions and human preferences, teaching the model to follow instructions and perform specific tasks. However, for a 1.6 trillion-parameter MoE architecture model, full-parameter post-training still places very stringent requirements on underlying hardware such as GPU memory capacity, inter-card communication bandwidth (such as the all-to-all communication triggered by MoE routing), and the stability of large-scale clusters. Leveraging the Huawei Ascend 910C computing power cluster with over a thousand chips, the joint research team successfully overcame communication bottlenecks through optimized distributed load-bearing and load balancing strategies. During the training process lasting over 1500 steps, the system did not encounter any interruptions, the model's computing power utilization rate (MFU) exceeded 30%, the key operator efficiency improved by 14%, and all indicators met industrial operational standards.
Industry analysis points out that the successful operation of the Huawei Ascend 910C cluster in training trillion-scale models confirms the technical feasibility of domestic AI chips in handling ultra-large-scale model deep training tasks. Prior to this, the development of large models heavily relied on NVIDIA GPU clusters, and domestic computing power mainly supported inference or small parameter fine-tuning tasks. The success of this joint research marks the acceleration of the domestic computing power ecosystem's transition from "inference-only support" to "carrying out full-parameter training of ultra-large models."
