Chinese Open Models Lead U.S. in Capabilities, Downloads, Academic Use

Chinese open-weight AI models hold a clear lead over American counterparts across multiple measures. A briefing prepared for U.S. Congressional members finds that Chinese models took the Hugging Face download lead in July 2025 and hold roughly a 1.6 billion download advantage. On the Artificial Analysis Intelligence Index, the top three Chinese models score 42-45 while leading American open-weight models score 23-26, with 15 Chinese models ranking above the top American open models. In academic publishing, Chinese open-weight models appear in over 40% of AI/ML papers on arXiv versus about 30% for U.S. models; Alibaba's Qwen is cited in 30% of papers compared to Meta's Llama at 21%. The briefing estimates that distillation of American frontier models accounts for only 1-2 months of the performance gap, suggesting Chinese lab strength is largely independent of that technique. Separately, analyst Rohan Paul argues open models account for roughly 78% of AI usage versus 21.6% for closed models, and that open models do not need to match frontier capabilities to pressure closed labs, only to become the default for the majority of use cases.