Huawei has successfully completed the post-training of DeepSeek V4-Pro, a 1.6 trillion parameter AI model, using a cluster of at least 1,000 Huawei Ascend 910C chips. The research was carried out collaboratively by Huawei, the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data.
The achievement carries weight beyond its technical specifications. Post-training is the process that transforms a raw language model into one that follows human instructions, aligns to safety guidelines, and behaves usefully in real-world applications. It requires sustained, high-intensity computation that is fundamentally more demanding than inference workloads.
Why training is harder than inference
Until recently, Chinese AI development had a clear weakness: Chinese chipmakers could supply hardware capable of running existing models (inference), but the more demanding training process, particularly the advanced post-training phase, remained largely dependent on Nvidia hardware subject to US export controls.
Training requires chips to coordinate constant bidirectional communication across thousands of cores simultaneously. Researchers involved in the project described the workload as involving complexity comparable to complex flyovers and loops, multiplying computational and communication demands several times over compared to inference tasks.
Completing post-training of a 1.6 trillion parameter model on domestic chips demonstrates that Huawei’s Ascend architecture has reached a threshold of practical training capability. The project used full-parameter refinement, meaning the entire model architecture was updated during the process rather than only targeted layers.
The geopolitical context
The US government has tightened export controls on advanced semiconductors to China multiple times since 2022, with the explicit goal of limiting China’s ability to develop frontier AI systems. The Huawei-DeepSeek collaboration is a direct response to those restrictions, demonstrating that Chinese institutions are finding ways to close the capability gap using domestically produced hardware.
Officials involved in the project stated that this work will help enhance the self-reliance of China’s AI industry chain. DeepSeek V4-Pro’s 1.6 trillion parameters make it the largest model DeepSeek has released.
Implications for the global AI industry
For AI developers and enterprises outside China, this development signals that the global AI landscape is becoming more multipolar. Chinese labs are moving up the capability stack from inference to training, and the hardware ecosystem enabling that shift is maturing.
For businesses in Latin America evaluating AI vendors and infrastructure partners, the emergence of competitive Chinese AI models trained entirely on domestic hardware introduces additional options as the market diversifies.
For sales teams tired of cold leads, slow customer responses, and manual processes, Dapta is the ultimate tool.
Dapta is the leading platform for creating AI sales agents specifically designed to increase inbound lead conversion. Respond to your leads in less than a minute with voice AI and WhatsApp that converts.
If you want your team to sell more while AI handles the complex stuff, you have to try it.
The successful training of DeepSeek V4-Pro on Ascend chips is a concrete step toward a world where advanced AI development is no longer bottlenecked by access to US-manufactured semiconductors. For the global AI industry, it signals that the competitive landscape is expanding geographically as well as technically.
