Huawei's Ascend Chips Successfully Enable DeepSeek-V4-Pro Post-Training, Advancing China's AI Self-Reliance
Key Takeaways
- ▸Huawei's Ascend 910C chips successfully powered full-parameter post-training of DeepSeek-V4-Pro (1.6 trillion parameters) on a cluster of at least 1,000 chips
- ▸Post-training—the complex process of teaching models to follow instructions and safety rules—demands several times more computational resources than inference, making this a significant technical breakthrough
- ▸This achievement marks major progress toward China's AI semiconductor independence amid US export restrictions on advanced AI chips
Summary
A research consortium including Huawei Technologies announced that it successfully used Huawei's Ascend 910C chips to complete post-training for DeepSeek-V4-Pro, the company's largest model to date with 1.6 trillion parameters. The project ran the model on a computing cluster powered by at least 1,000 Huawei chips and achieved full-parameter post-training, meaning the entire model architecture was refined without compromises. This represents a major technical achievement in China's efforts to develop independent AI semiconductor capabilities amid increasing US sanctions on chip exports.
Post-training—where models learn to follow human instructions, safety rules, and specific tasks—is significantly more computationally complex than inference, the process of running a finished model to answer user prompts. While Chinese chipmakers have successfully supported inference tasks, achieving post-training capability on domestic hardware is a far more challenging undertaking. The researchers described the leap from inference to post-training as transforming a simple one-way road into a complex system with "flyovers and loops," multiplying computational and communication demands several times over.
The collaborative project, conducted by Huawei, the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data, underscores China's growing capability to support advanced AI model development using domestic hardware. This achievement is significant in the context of Beijing's push for technological self-reliance in AI, particularly as Western nations tighten restrictions on semiconductor access.
Editorial Opinion
This breakthrough is genuinely significant for China's AI infrastructure independence, demonstrating that domestic chips can now handle the computationally intensive phase of large-scale model refinement. However, it's worth noting that post-training is still just one phase of the AI development pipeline—the ability to conduct large-scale pre-training (the even more computationally expensive phase of initially training from scratch) remains the next frontier for Chinese semiconductor capabilities under current export restrictions.
