ATOM Report: China Passed the US in Open-Weight AI Downloads -- 1.15B to 723M, Qwen in the Lead
The ATOM Report, measuring the open language model ecosystem across tracked downloads, finds China surpassed the United States in open-weight model downloads in summer 2025. As of March 2026 the gap has widened: 1.15 billion tracked downloads for Chinese models versus 723 million for US models.
How the Gap Formed
Qwen is the primary driver. The model family became the default base for a large portion of fine-tuners, application builders, and researchers who want models they can run and modify cheaply. Qwen’s dominance is not built on one flagship release but on breadth: useful models across many sizes, with a particular strength in the small and mid-size range that practitioners can deploy on accessible hardware.
DeepSeek occupies a different position. It does not dominate the full open-model ecosystem the way Qwen does, but it leads the large-model category, specifically models above 250 billion parameters. For the subset of practitioners working at the highest capability tier with open weights, DeepSeek is the reference point.
US Models Still Competitive on Momentum
After adjusting for model size and age, the picture is more nuanced. US models including GPT-OSS 120B and Nemotron Super 120B show strong adoption momentum even as the raw download count favors China. The overall aggregate is tilted by the sheer number of Qwen variants and the volume of downloads at smaller sizes, not by a uniform Chinese lead across every capability tier.
What the Download Gap Means
Download metrics capture something leaderboard rankings do not: what practitioners actually use when they own the weights. A model that tops the Artificial Analysis Intelligence Index gets covered heavily. A model that becomes the default fine-tuning base for thousands of applications shapes the actual deployed software stack without necessarily holding the headline ranking.
China’s lead in open-weight downloads reflects Qwen and DeepSeek successfully targeting the builder layer rather than the benchmark layer. For the organizations building products on open weights, the infrastructure underneath is increasingly Chinese.
The ATOM findings also have a policy dimension. US export controls on frontier chips are designed to limit Chinese model training capability. They have not prevented Chinese models from dominating the global open-weight distribution layer. Capability and distribution have partially decoupled.