Chinese models with open weights have surpassed American ones in popularity

Analysis of the market of open models showed a significant advantage of Chinese developments over American counterparts in terms of downloads and testing results. According to the AAII benchmark data, leading models from China score from 42 to 45 points, while similar solutions from the US demonstrate performance in the range of 23–26 points.

The study highlights that the dominance of Chinese companies is observed in the segment of models with open weights (open-weight). At the same time, American developers maintain their leadership in the category of fully open-source solutions, where not only the executable code but also the complete training history with datasets is available. The gap in AAII scores indicates a more aggressive strategy for distributing and optimizing Chinese neural networks for end-users.

Download statistics confirm the shift in market balance: users more often choose Chinese architectures for local deployment. Despite high scores in tests, experts note the difference in approaches to licensing and development transparency between the two regions. The full report and discussion of the evaluation methodology are available on the Hacker News platform.

These results change the perception of competition in the field of generative AI, where American laboratories previously dominated. The growing influence of Chinese models with open weights creates new conditions for developers, who now more often prefer more productive but less transparent systems in terms of training.