A collaborative effort involving Shenzhen Riverfront College, Harbin Institute of Technology (Shenzhen), and Huawei has successfully completed full-parameter post-training for the DeepSeek-V4-Pro model on a domestic computing cluster. This marks a significant advancement, demonstrating that China's AI infrastructure can now handle the training of ultra-large models, not just inference and fine-tuning. The project achieved stable training of over 1500 steps with a model compute utilization (MFU) exceeding 30% on the Ascend 910C cluster, showcasing breakthroughs in distributed training, sparse model optimization, and long-term training stability. This achievement not only validates the capability of domestic AI hardware but also serves as a practical training ground for cultivating AI talent skilled in large-scale model development. AI
IMPACT Demonstrates the growing capability of domestic AI infrastructure to train frontier models, potentially reducing reliance on foreign hardware and fostering local AI development.
RANK_REASON This article details a research and development achievement in training a large AI model using domestic hardware, rather than a new model release or product launch by a major AI lab. [lever_c_demoted from research: ic=1 ai=1.0]
- Ascend 910C
- Deep City AI Computing Platform
- DeepSeek-V4-Flash
- DeepSeek-V4-Pro
- Harbin Institute of Technology (Shenzhen)
- Huawei
- Huawei 2012 Laboratory
- Huawei Computing Product Line
- Huawei GTS
- Shenzhen Big Data Research Institute
- Shenzhen Riverfront College
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
