The current balance of power in open models

Key Points:

• Open-weight language models, whose weights are publicly available, differ from closed models accessed only via APIs; Chinese AI companies have led open-weight models since April 2025, while true open-source models (including training code and data) are primarily developed by U.S. non-profits.
• Chinese open-weight models have surpassed American counterparts in downloads and benchmark performance, with models like GLM-5.3 and Kimi K3 outperforming U.S. models by significant margins; the performance gap to leading closed American models is estimated at 2-5 months.
• Chinese labs’ faster release cycles and focused task distribution contribute to their competitive edge, alongside strategic purchases of cutting-edge training data; however, distillation from American models explains only a small part of their success.
• Adoption of Chinese open-weight models is growing rapidly in academia and industry, with Chinese models like Alibaba’s Qwen mentioned in over 40% of recent AI research papers and widely used by startups and major companies, indicating a shift in global AI research leadership.
• The rise of open-weight models introduces new risks, such as cybersecurity threats, and regulatory uncertainty; experts recommend continued U.S. investment in open models to maintain influence, manage risks, and foster AI diffusion in the domestic economy.