State of Open Models: Summer 2026 Observations

Hugging Face’s summer read on open models lands in a weird place: everything is growing, but the growth is lopsided enough to make the headline numbers almost misleading. From January to August 2026, public model repos on the Hub rose from 2.43 million to 2.96 million, datasets from 711,000 to 1 million, and Spaces from 1.00 million to 1.44 million. Yet the usual power-law holds hard. About 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.

The most striking change is at the frontier. Several Chinese labs skipped the old ritual of starting small and working upward. In most months of 2026, the largest open model from a Chinese lab was bigger than anything released by an American lab. China’s monthly peak ranged from 754B to 2.78T parameters. The U.S. stayed under 130B in five of seven months, with NVIDIA’s Nemotron 3 Ultra at 561B in May and June and Thinking Machines Lab’s Inkling also in the mix. Some Chinese labs now ship almost nothing below 70B, which means the first thing many developers see is a model they can’t run as-is.

But downloads and likes are telling different stories. Hugging Face compared the top 25 model repos by downloads this year with the top 25 by likes, and only one repo showed up on both lists. None of the 2026 releases made the download top 25; thirteen of the 25 dated from 2022. all-MiniLM-L6-v2 was downloaded 1.55 billion times in seven months but got just 5,156 likes. Kimi-K3, by contrast, was downloaded about 60 times for every like. Likes track excitement. Downloads track what people actually wire into work.

That same split shows up in licensing and in who is building on what. Among 178 Chinese releases above 20B parameters this year, 59% use Apache 2.0 and 22% use MIT, with no non-commercial restrictions at all. DeepSeek and Z.ai are shipping models between 700B and 1.65T under plain MIT. So the money is clearly not in licence fees. It’s in APIs, clouds, hardware, and ecosystem pull.

And that ecosystem pull now has a name: Qwen. Hugging Face says Qwen-based models account for 151,448 derivatives on the Hub, 2.6 times Meta’s total and 4.7 times Llama’s repositories specifically. Those derivatives have been growing by roughly 180 to 210 new repos a day through the first seven months of 2026. Small models still dominate actual usage, though. Models under 1B get 83% of all-time downloads, and even in 2026 models above 70B account for just 3% of volume. The frontier gets the headlines; the boring little models keep the lights on.