A high-level research consortium in Shenzhen has successfully completed the full-parameter post-training of a 1.6-trillion-parameter AI model, DeepSeek-V4-Pro, utilizing entirely domestic hardware. This achievement, spearheaded by the Hetao Institute of AI in collaboration with Huawei and the Harbin Institute of Technology (Shenzhen), represents a critical milestone in China’s quest for technological self-reliance. By leveraging a domestic computing cluster powered by Huawei’s Ascend 910C chips, the project demonstrates that Chinese silicon can now support the most demanding tier of artificial intelligence development.
The scale of this success is found in the technical distinction between model inference and full-parameter training. While many organizations can run existing models on diverse hardware, the intensive process of training a 1.6-trillion-parameter architecture requires massive computational throughput and seamless interconnectivity—areas where Western sanctions on high-end Nvidia GPUs were designed to cripple Chinese progress. The successful completion of this trial suggests that the domestic ecosystem is maturing rapidly enough to bypass traditional hardware bottlenecks.
Shenzhen’s role as the vanguard of this movement is no accident. The project integrated resources from the Shenzhen Big Data Research Institute and the Deep Shenzhen AI computing platform, showcasing a highly coordinated effort between municipal government, academia, and industry. This 'Shenzhen Model' of innovation is designed to mitigate the risks posed by volatile global supply chains by creating a closed-loop environment for AI development, from chip design to model deployment.
While the 1.6-trillion-parameter count suggests a Mixture-of-Experts (MoE) architecture similar to leading global models, the true significance lies in the validation of the 'domestic path.' For Chinese enterprises, the fear of being locked out of the generative AI revolution is being replaced by a cautious optimism that local clusters can finally bear the weight of frontier-level research. This development effectively serves as a proof of concept for a parallel AI infrastructure that operates independently of Silicon Valley’s hardware stack.