Chinese laboratories have strengthened their lead over US developers in open-weight language models, according to an analysis prepared for a US congressional briefing. Open-weight models publish the parameters needed to run and adapt them, although they do not necessarily release the training data and code required for full reproducibility.
The analysis says Chinese models have recorded about 3.2 billion Hugging Face downloads, roughly twice the US total. It also points to a clear performance gap on the Artificial Analysis Intelligence Index: the leading Chinese entries scored between 42 and 45, while the top American open models scored between 23 and 26. The author estimates that Chinese open models trail the strongest closed US systems by two to five months, compared with six to nine months for American open models.
That advantage is especially visible in agentic coding, where models carry out multi-step software tasks. Faster release cycles and a narrower focus on high-demand tasks partly explain the results, but the analysis argues that the models represent genuine competition rather than benchmark optimization alone.
The figures are an analyst’s estimates, not an official government assessment. They nevertheless underline a strategic divide: US companies still lead at the closed frontier, while Chinese labs increasingly supply the models that outside developers can download and deploy themselves.