Hugging Face has published its State of Open Models: Summer 2026 report, detailing seven months of platform metrics that highlight a widening divergence between benchmark attention and production deployment.
According to the analysis covering January through August 2026, public model repositories on the Hub grew from 2.43 million to 2.96 million, public datasets surpassed 1 million for the first time, and Spaces expanded to 1.44 million. Despite the catalog expansion, usage remains heavily concentrated: 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all platform download volume.
Attention Versus Production Adoption
A key finding of the report is the structural separation between community interest and production integration. Comparing the top 25 model repositories by 2026 downloads against the top 25 by likes, exactly one repository appeared on both lists.
Thirteen of the top 25 downloaded models in 2026 were originally released in 2022 or earlier. For instance, the 22.7M-parameter embedding model sentence-transformers/all-MiniLM-L6-v2 recorded 1.55 billion downloads over seven months while holding just 5,156 likes. Conversely, frontier releases like Moonshot's 2.8T-parameter Kimi-K3 accumulated tens of thousands of likes while recording roughly 60 downloads per like.
Hardware constraints continue to dictate deployment volume. Among models with declared parameter counts, architectures under 1 billion parameters capture 83% of all-time downloads. Models exceeding 70 billion parameters accounted for only 3% of total downloads accumulated during the first seven months of 2026.
Qwen Leads Open Model Derivatives
Alibaba's Qwen architecture has established a decisive lead in open-weight fine-tuning and downstream packaging. Qwen-derived repositories reached 151,448 on the Hub, representing 2.6 times Meta's total derivative footprint and 4.7 times the number of Llama-based repositories. New Qwen derivatives have expanded at a rate of 180 to 210 repositories per day throughout 2026.
Qwen's multi-tier release strategy across model scales drove 2.05 billion downloads across declared parameter repositories in 2026, roughly 55 times the 37 million downloads logged by Moonshot's frontier-only catalog.
In local inference formats, Qwen logged 39.6 million monthly GGUF downloads, outpacing Google's Gemma (20.8 million) and Meta's Llama (7.5 million). Growth across execution runtimes substantially outpaced core model hosting: repositories specifying the gguf library expanded by 464%, robotics framework lerobot grew 194%, and Apple's mlx framework grew 148%, compared to 16% growth for transformers and peft.
Licensing Dynamics Across Frontier Weights
Chinese research organizations released models with parameter counts ranging between 754 billion and 2.78 trillion throughout 2026. Of 178 Chinese model releases above 20 billion parameters, 59% were published under Apache 2.0 and 22% under MIT licenses.
On the American side within the same size band, 29% used Apache 2.0 or MIT, 41% used custom licenses, and 30% did not declare a license. Hardware manufacturers NVIDIA and AMD led U.S. open-weight publications, releasing over 200 model repositories each to support hardware-optimized runtimes.
Autonomous Agents Emerge as Primary Hub Users
Hugging Face's newly published agent-usage dataset, which monitors automated calls from coding assistants and agent runtimes via huggingface_hub and the hf CLI, revealed that autonomous agents have become major consumers of the platform.
In July 2026, Anthropic's Claude Code accounted for 44.4% of identified agent traffic (down from 67.8% in April), while OpenAI Codex grew its share from 10.4% to 20.8%. Unregistered client harnesses accounted for 24.3% of agent requests in July and 59.8% in May, indicating rapid turnover in agent development frameworks.
