Back to Stories

Chinese developers account for over 45% of top open-model public downloads



Robert PraasPierre-Alexandre BallandFrancisco Ríos
December 17, 2025 - 2 min read

2025 was a pivotal year for open source models, surpassing the 2 million mark and seeing robotics and multimodal tasks flourish. From recently uploaded models, Qwen, DeepSeek and Llama were downloaded the most, based on downloads in 2025, counting downloads up to one year after each model’s release. Most notably, over 45% of these downloads came from Chinese models.

Qwen2.5 models, ranging from 0.6 to 7B in size, were collectively downloaded more than 750 million times this year. Despite being released in 2024, Meta Llama’s 3.1 and 3.2 models were also downloaded more than 500 million times this year. The data comes from the Open Model Evolution dashboard, accompanying the recent Economies of Open Intelligence paper, opening up a large dataset of Hugging Face data.

Beyond text models, video and visual creation tooling drives adoption, exemplified by ByteDance’s AnimateDiff-Lightning, the Comfy/ComfyUI ecosystem as well as releases from Stable Diffusion and Black Forest Labs.

Despite the abundant influence of Chinese and US developers, significant activity is user or community-driven: Hexgrad, Bartowski, the MLX community and many more creators contributed with heavily adopted models.

Importantly, the public data includes the multiple downloads of a user per model. This can happen because a user re-downloads a model to re-use it, or they actively download a model multiple times to boost the download counts as performance metric. The Economies of Open Intelligence paper provides a valuable insight into deduplicated downloads: the table below shows for models, developers and countries, both for all-time and a one-year time frame, which entities received the most downloads from unique users (and hence, only one download per user is counted). When accounting for unique downloads, larger models like DeepSeek-R1 turn out to be very popular. Similarly, the LM Studio Community's models were downloaded the most in one year's time (noting that their models only have to be downloaded once to be further used in their software).


Source: Economies of Open Intelligence




Scan the QR code to view this story on your mobile device.


open modelsdownloadshugging face