A New Benchmark for Domestic Computing Power
In the pursuit of technological self-reliance, computing power independence stands as a critical frontier. Exciting news has emerged from Shenzhen, where a collaborative research team has achieved a formidable milestone: successfully conducting full-parameter post-training for the colossal DeepSeek-V4-Pro AI model, which boasts a staggering 1.6 trillion parameters, utilizing China's homegrown computing infrastructure. This accomplishment represents far more than a technical feat; it is a comprehensive stress test of the nation's independent computing capabilities.
Synergistic Efforts: Uniting Academia and Industry
This breakthrough is the fruit of deep collaboration between academia and industry. The project brought together technical experts from a Shenzhen-based innovative research institute, the Harbin Institute of Technology (Shenzhen), the Shenzhen Big Data Research Institute, and specialist teams from a leading technology enterprise. Leveraging an AI computing platform, they jointly tackled the core challenge of training large-scale models on domestically developed hardware.
The Hardware Backbone: Ascend Clusters Rise to the Challenge
The entire training operation was powered by the domestic Ascend 910C AI computing clusters. Driven by this "homegrown core" system, the team adeptly managed the training of the massive DeepSeek-V4-Pro model. It is crucial to distinguish between using a pre-trained model for inference and performing full-parameter training—the latter imposes extreme demands on hardware stability, parallel efficiency, communication bandwidth, and ecosystem compatibility. This success serves as the most substantial validation yet for China's indigenous computing power.
Far-Reaching Implications: Feasibility Proven in Practice
The significance of this endeavor extends well beyond a single training run. Firstly, it provides a valuable precedent and积累了经验 for third-party research institutions worldwide, demonstrating that world-class model training can be accomplished on computing platforms not reliant on mainstream international chips. Secondly, it offers robust evidence that China's AI chip design and manufacturing prowess now meets the requirements for supporting cutting-edge AI research and industrialization. The feasibility of the related hardware and software technology pathway has been empirically verified, laying a solid foundation for greater autonomy in China's AI infrastructure.
Looking Ahead: An Independent Ecosystem Takes Shape
This achievement is a clear signal that domestic computing power is transitioning from being "usable" to becoming "efficient, reliable, and competitive." It not only boosts industry confidence but will also attract more developers and enterprises to build AI ecosystems based on domestic hardware. As such practices multiply and the technology chain matures, a healthy, vibrant, and self-sufficient AI computing ecosystem is rapidly taking shape in China, promising a solid and dependable foundation for the future digital economy.