State of Open Models: Summer 2026 Observations

Hugging Face published its State of Open Models: Summer 2026 Observations, covering activity on the Hugging Face Hub from January to August 2026. The report says public model repositories grew 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. It states that roughly 85.6% of models have fewer than 200 lifetime downloads and that 1.5% of repositories account for 99.2% of all downloads.

On model scale, the report says that in almost every month of 2026 the largest and most performant open model from a Chinese lab was larger than any model an American lab released, with China's monthly ceiling running between 754B and 2.78 trillion parameters, while U.S. models stayed under 130B in five of seven months. It cites NVIDIA's Nemotron 3 Ultra at 561B in May and June, and Thinking Machines Lab's Inkling, as exceptions. AMD and NVIDIA each released more than 200 new model repositories, with LiquidAI third at around 100, according to the report.

The report says downloading and liking measure different things: of the top 25 repositories by downloads and by likes, only one appears in both lists, and thirteen of the 25 download leaders date from 2022. It states 59% of 178 Chinese releases above 20B parameters carry Apache 2.0 and 22% MIT, and that Kimi K3 and Qwen 3.8 2.4T recently began including non-commercial restrictions and revenue share requirements. A comment on the article disputes the claim that no Chinese releases carry non-commercial restrictions.

The report says Qwen-based models account for 151,448 derivatives on the Hub, 2.6 times Meta's total footprint, with Google at 82,506. It also states that models under 1B take 83% of all-time downloads, that GGUF, lerobot and mlx repositories grew 464%, 194% and 148% respectively, and that Claude Code led July agent traffic at 44.4% while Codex climbed from 10.4% to 20.8%. It notes a documented autonomous agent intrusion in July, and that analysis of captured attack code was completed on a quantized GLM-5.2 model after frontier closed models' guardrails declined the work.