OpenRouter’s latest weekly call-volume chart put five Chinese-developed models in the top five, and on Hugging Face, China’s open-source models lead cumulative downloads, with the Qwen family past 1 billion downloads at number one. The signal is not one model’s benchmark but the breadth of global usage now running on Chinese weights.
Calls and downloads
‘Calls’ mean individuals who lack local deployment pay per token to use a model online. ‘Downloads’ are organisations, banks, governments and energy firms pulling models in to embed in their own systems, the core ‘AI plus’ indicator, and China leads there as well. Across Africa, with 1,500 to 3,000 languages, many local-language training efforts now start from Chinese models. US investment data shows as many as 80 per cent of American AI startups use a Chinese open model in their fundraising pitch, and more than 200 derivative models appear worldwide each day.
Why the world is downloading
On 27 July, Moonshot open-weighted Kimi K3 at 2.8 trillion parameters, the largest open model to date, and it reached number one on Hugging Face’s overall trending chart within an hour. Researchers say the scale of open weights shows Chinese firms can train super-large models, optimise sparse architectures, run low-cost inference and adapt globally. Experts frame continuous open-sourcing as trading model investment for long-term ecosystem influence, with revenue later from cloud inference, fine-tuning, agent platforms and industry solutions. China’s models are moving from single-point breakthroughs to coordinated technology, industry and ecosystem development, though the gap from usage lead to full technology lead still needs work on compute autonomy and foundational ecosystems.
Read the original report (Sohu IT)
Translated and adapted from Sohu IT (it.sohu.com).