A research collaboration featuring Huawei Technologies has successfully used domestic Ascend 910C chips to complete post-training of the DeepSeek-V4-Pro model, a crucial step enhancing China’s capacity to independently develop advanced AI amid mounting US sanctions on semiconductor exports.
- Huawei chips complete complex AI model post-training for the first time.
- Full-parameter training performed on DeepSeek-V4-Pro with 1.6 trillion parameters.
- Move reduces China’s reliance on restricted US semiconductor technology.
What happened
A joint research team including Huawei Technologies has achieved a significant milestone by using Huawei’s Ascend 910C chips to complete post-training of the DeepSeek-V4-Pro AI model. This process involved running the largest DeepSeek model to date, consisting of 1.6 trillion parameters, on a large computing cluster powered by more than 1,000 Huawei chips. The team successfully updated and refined the entire model architecture through full-parameter post-training, an advancement beyond earlier applications which mainly focused on AI inference.
This accomplishment marks a transition in China's semiconductor use for AI—from merely running pre-existing models to undertaking the more computationally intensive process of training. The cluster managed to execute over 1,500 uninterrupted training iterations, demonstrating system stability and effectiveness, and improving the model’s mathematical capacity. The collaboration also included institutes based in Shenzhen and the Harbin Institute of Technology.
Why it matters
China’s AI industry has been constrained by US export controls limiting access to advanced training hardware from American chip manufacturers such as Nvidia and AMD. While Chinese companies have made strides in AI inference workloads, full-scale training capability has remained elusive due to the complexity and computational demands involved. Achieving full-parameter post-training domestically signifies a major step towards technological self-reliance in AI development.
Training AI models requires far more sophisticated hardware and infrastructure than inference, often necessitating months of compute time and massive clusters. The success with Huawei’s chips reduces dependency on foreign hardware in the face of escalating trade restrictions and sanctions, enabling China to maintain competitiveness in large-scale AI model development. This also opens prospects for accelerating innovation in AI sectors and ensuring continuity amid geopolitical uncertainties.
What to watch next
Going forward, attention will focus on how quickly China can scale up this post-training capability and whether domestic chipmakers can close the gap in full pre-training from scratch. Other major Chinese AI firms, including Baidu and Meituan, are also exploring training on domestic chips, signaling increased momentum in indigenous AI hardware development. Huawei’s continued efforts with advanced AI architectures and infrastructure, such as its new “Agentic Infra” paradigm, indicate an expanding ecosystem beyond chatbot applications.