China's Open-Weight AI Models Trail US Closed Models by Seven Months
An analysis of an open-source AI benchmark dataset called whatmodel.sh reveals that Chinese open-weight models currently trail US closed models by a margin of approximately seven months. This data-driven insight provides a transparent, quantifiable measure of the global AI race, helping developers and industry observers understand the performance gap between open and proprietary ecosystems. The underlying dataset incorporates 18,755 text model results across 1,211 models and 617 benchmarks, combining both independent evaluations from academic papers and lab-reported scores.
## BACKGROUND
Open-weight models provide users with downloadable model parameters that can be hosted locally and audited, whereas closed models are typically accessed exclusively via proprietary APIs. Evaluation frameworks like EleutherAI harness and EvalPlus allow researchers to systematically test and benchmark these large language models across a wide array of standardized tasks.