Chinese AI labs lead in open-source model scale, Hugging Face report

A report from Hugging Face titled ‘State of Open Models: Summer 2026 Observations’ reveals a significant shift in the artificial intelligence landscape, highlighting the rise of Chinese open-source models. Data from January to July 2026 shows that Chinese laboratories frequently released the largest open models, with parameter counts reaching between 754 billion and 2.78 trillion. In contrast, US-based labs typically released models under 130 billion parameters during most months of the period.

Alibaba’s Qwen model has emerged as a dominant community standard. It has seen over 151,000 derivative models created, which is 2.6 times the total footprint of Meta and 4.7 times the number of derivatives for Llama. Qwen’s download volume is significantly higher than its competitors, driven by a strategy of providing a wide range of model sizes and utilizing permissive licenses like Apache 2.0.

While large-scale models are gaining prominence, small-scale models remain the industry workhorse. Models with fewer than 1 billion parameters account for 83% of all-time downloads, whereas models exceeding 100 billion parameters represent only 1%. Additionally, the report notes that hardware vendors like NVIDIA and AMD are among the most frequent publishers of new open models, with both companies releasing over 200 new models during the period.