Huawei's Ascend 910C chips complete post-training on DeepSeek's 1.6T-parameter model

A research collaboration involving Huawei Technologies, the Shenzhen Loop Area Institute, Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data has successfully used Huawei's Ascend 910C chips to complete full-parameter post-training on the DeepSeek-V4-Pro model. This 1.6-trillion-parameter model is the largest DeepSeek has built, and the team deployed a computing cluster with at least 1,000 Huawei chips to run the process. The work represents a significant milestone for China's semiconductor industry, which has been constrained by US export controls and previously focused mainly on AI inference (running finished models) rather than the far more complex training stage.

Post-training teaches a model how to follow instructions, apply safety rules, and perform specific tasks—analogous to building complex flyovers and loops on a previously one-way road. The researchers achieved "full-parameter" post-training, meaning the entire architecture was updated without shortcuts, enabling the model to self-reflect and adjust. The Shenzhen government’s social media post highlighted that this project "will help enhance the self-reliance of China's AI industry chain." While Chinese chips have made strides in inference, this demonstration of domestic chips handling high-end training workloads could shift the AI hardware landscape and reduce dependence on foreign GPU suppliers.