The same dashboard can tell two different stories

A new open model can dominate a week of conversation and still barely appear in the systems that people run every day. A small embedding model can attract almost no excitement and quietly sit inside thousands of retrieval pipelines. Both are forms of success, but they are not the same form.

Hugging Face’s summer analysis of its own Hub is useful because it keeps those signals apart. It looks at models, datasets, Spaces, downloads, licences, derivatives and a newer slice of traffic from coding agents. The result is less a ranking of the best model than a picture of what a large open-model platform is being used for.

The most important caveat arrives with the data itself. Hub activity is not the entire AI market. It does not capture private deployments, API calls elsewhere or models distributed through other channels. Hugging Face says so directly. That makes the report a strong view of one ecosystem, not a universal scoreboard.

Attention does not equal adoption

The report says public model repositories on the Hub rose from 2.43 million to 2.96 million in the first seven months of 2026. Datasets grew from 711,000 to one million and Spaces from one million to 1.44 million. Under that headline growth sits a very uneven distribution: according to Hugging Face, 85.6% of models had fewer than 200 lifetime downloads, while 1.5% of repositories accounted for 99.2% of downloads.

The analysis contrasts likes with downloads. Likes tend to show what the community is noticing around a release. Downloads are more likely to reflect a model wired into a repeated task. Hugging Face says no model published in 2026 reached its top 25 by downloads in the measurement window, while many long-running models did.

That is a reminder, not a verdict on quality. Older models have time to become infrastructure, and downloads can be automated. But the distinction is practical for anyone choosing a model: novelty is not the same as a dependable deployment history.

Where open-model ecosystems accumulate

Hugging Face reports that Qwen-based models had 151,448 derivatives on the Hub, compared with a smaller footprint for other large model families. A derivative can be a fine-tune, conversion or packaging change. It is evidence that people are building on a base model, not proof that every derivative is active, good or commercially successful.

The report also argues that the licence choices of large open models are part of the competitive picture. It counts a high share of permissive terms among the Chinese releases it examined, while warning elsewhere that terms can change and licences are not all alike. The value of an open release may come through cloud demand, hardware, services or ecosystem position rather than licence fees.

That is an interpretation from Hugging Face’s analysis, not a disclosed business model for every lab. Still, the basic point holds: an open-weight release can be less about selling a file and more about becoming the default material people build with.

Agents are now visible in the traffic

The report also uses a new Hugging Face dataset that records the share of agent-attributed requests made through the Hub library and command-line tools. A request only appears when the client includes a recognised agent identifier. It is a partial but unusually concrete window into how coding tools touch an open-model platform.

For July, the dataset lists Claude Code with 44.4% of named agent-attributed requests and Codex with 20.8%. It also records an unknown category at 23.1%. The shares are not market shares for all coding agents. They are shares of this labelled slice of Hub traffic.

That limitation is exactly why the data is useful. It does not try to claim more than it can see. It shows a changing layer of automated users searching for models, pushing data, running jobs and creating Spaces. The people who browse a model page are no longer the only audience for the Hub.

What is confirmed, what Hugging Face infers, and what is open

Confirmed: Hugging Face published the analysis on 14 August 2026, based on Hub activity from the first seven months of the year. It published the underlying agent-usage dataset separately and describes the request-labelling method used there.

Hugging Face’s interpretation: attention and adoption are different signals, Qwen has become a central base-model ecosystem on the Hub, smaller models remain the practical layer and agents are becoming meaningful users of the platform.

Open questions: how well these patterns match private deployments and other distribution channels, how comparable model downloads are across different tasks, whether the agent labels are complete, and how far an ecosystem signal can tell us about commercial success or model quality.

Sources

  1. Hugging Face — State of Open Models: Summer 2026 ObservationsPrimary Hugging Face analysis, published 14 August 2026. Source for Hub growth, download concentration, derivatives, licensing analysis and the report’s methodological limits.
  2. Hugging Face — Agent Usage on the Hugging Face Hub datasetPrimary public dataset. Source for the labelled agent-attributed request shares and the explanation of what the dataset can and cannot measure.