63% of new AI models are now based on Chinese tech
Alibaba released its Qwen series as an open-weight model family at the Consumer Electronics Show (CES) in Las Vegas this week, a move that has already spurred the creation of over 100,000 derivative models on platforms like Hugging Face. This development highlights a major push by Chinese technology companies to lead in open-source artificial intelligence, with more than 1,000 Chinese firms participating in the event which concluded on January 10. Alongside Alibaba, Chinese research lab DeepSeek introduced its R3 reasoning model under an open license, further emphasizing the shift toward collaborative development.
Data from a Stanford analysis reveals that by September 2025, Alibaba’s Qwen family had surpassed Meta’s Llama as the most downloaded language model family on Hugging Face. By that time, 63 percent of all new fine-tuned models were based on Chinese foundation models. This “open-weight” approach allows developers to modify trained parameters even if the training code remains proprietary, a method now being adopted by robotics companies like Unitree and Agibot for complex tasks in industrial and home settings.
Nvidia CEO Jensen Huang supported this trend during his keynote, stating that open innovation ensures AI technology proliferates without leaving anyone behind. The event also featured the CES Asia Night, where U.S. and Asian leaders discussed cross-border growth, including partnerships between Dolby Laboratories and Chinese manufacturers Hisense and TCL. Additionally, Nvidia contributed to the open ecosystem by announcing its Alpamayo models for autonomous vehicles and updates to its Cosmos foundation model for physical AI applications.