Hugging Face’s Summer 2026 report describes a rapidly expanding open-model ecosystem that remains unevenly used. Published on August 14, the report covers growth between January and August 2026 across public model repositories, datasets and Spaces, while showing that a small group of older projects still attracts most downloads. State of Open Models: Summer 2026 Observations
The Hub is getting larger, but attention is not spreading evenly
The number of public model repositories rose from 2.43 million to 2.96 million during the period covered. Public datasets increased from 711,000 to 1 million, and public Spaces grew from 1 million to 1.44 million. For researchers and developers, that means more open AI assets to inspect, adapt and deploy. The counts describe a broader supply, though they do not show that every new repository is attracting meaningful use.
Downloads remain concentrated in a narrow part of the Hub. About 85.6% of models had fewer than 200 cumulative downloads, while roughly 1.5% of repositories accounted for 99.2% of all downloads. No model published in 2026 appeared among the 25 repositories with the highest download totals; 13 of those 25 repositories dated from 2022.
The scale of the leading repositories also puts the concentration into perspective. The all-MiniLM-L6-v2 repository recorded 1.55 billion downloads in seven months and received 5,156 likes. A growing catalogue therefore does not automatically translate into a wider distribution of attention or adoption.
The largest open models are becoming more competitive
Hugging Face’s figures also show a widening gap at the top end of model size. The monthly ceiling for Chinese open models reached between 754 billion and 2.78 trillion parameters in 2026. US models stayed below 130 billion parameters in five of the seven months analyzed, although NVIDIA’s Nemotron 3 Ultra reached 561 billion parameters in May and June.
Corporate participation is visible in the repository data. AMD and NVIDIA each published more than 200 new model repositories during the year. Among 178 Chinese models above 20 billion parameters released in 2026, 59% used the Apache 2.0 license and 22% used the MIT license.
Those numbers connect model scale with company activity and licensing choices, but they do not establish that larger models are better adopted. The report’s download data points in the opposite direction for everyday usage: size leadership and practical visibility are not the same thing.
Qwen and smaller models remain central to actual usage
Hugging Face counted 151,448 repositories derived from Qwen models, with roughly 180 to 210 new Qwen-derived repositories appearing each day during the first seven months of 2026. That pace shows how quickly an established model family can become a base for further experimentation and releases.
Smaller models still dominate downloads. Models under 1 billion parameters represented 83% of cumulative downloads among repositories that reported their parameter count, while models above 100 billion represented 1%. The contrast limits any reading of parameter counts as a simple proxy for reach: the largest models attract attention in one part of the ecosystem, while smaller models account for most measured usage.
Agent tooling is becoming part of the platform
The report also tracks a shift in the tools appearing around open models. Claude Code’s share of identified agent traffic fell from 67.8% in April to 44.4% in July, while Codex rose from 10.4% to 20.8%. Nearly one quarter of July’s identified agent traffic came from frameworks that were not yet named in the dataset, compared with 59.8% in May.
Hugging Face added machine-readable Markdown for scientific papers in March, agent traces as a first-class dataset type and an agents.md endpoint for every Gradio Space in April, then added the hf_fs tool to its MCP server in July. These changes indicate that the Hub is being shaped not only as a catalogue for models and datasets, but also as infrastructure that agents can read and use.
Hugging Face’s report is therefore most useful as a snapshot of scale and direction. Open-model supply is growing, model families are generating large numbers of derivatives, and agent tooling is becoming more visible. Yet repository counts, parameter totals and downloads answer different questions; none of them alone provides a complete measure of model quality or adoption.