Hugging Face's State of Open Models: Summer 2026 report says the open-model ecosystem is expanding rapidly, but attention and adoption are pulling in different directions. The report's central finding is blunt: the models attracting the most likes are often not the models doing the most practical work.
Hugging Face Watch analysis
What happened
Published on 14 August, the report analyses activity on the Hugging Face Hub from January through August 2026. Public model repositories rose from 2.43 million to 2.96 million, while datasets and Spaces also grew sharply.
That abundance is highly concentrated. Hugging Face says 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads. The open ecosystem is busy. Its audience is choosy to the point of being statistically rude.
Key findings
- Likes are not adoption
Attention clusters around new frontier models, while downloads reflect older, smaller models wired into real pipelines. - Qwen is the community base layer
Qwen-based models account for 151,448 derivatives on the Hub, according to the report. - Small models still do the work
Models below 1B parameters account for 83% of all-time downloads among repositories declaring parameter counts. - The frontier is getting enormous
Chinese open-model releases reached monthly ceilings between 754B and 2.78T parameters during the period covered. - Open weights are a strategy
The report argues that permissive releases are more likely to support API, cloud, hardware or platform positioning than direct licence revenue.
Why it matters
The report offers a useful corrective to release-day coverage. A model can dominate social feeds and still be irrelevant to the software people run every day. Downloads, derivatives and the hardware needed to deploy a model tell a less glamorous but more durable story.
It also explains why Qwen's breadth matters. A family spanning small local models and much larger deployments gives developers a path to stay inside one ecosystem as their needs change, while downstream derivatives make that ecosystem harder to dislodge.
Our read
Treat frontier parameter counts as a measure of ambition, not automatic evidence of market traction. For practical AI, the strongest signal remains boringly operational: what developers can run, reuse and build on.
What to watch
- Whether Qwen's derivative lead continues through the next Hub snapshot.
- Whether large-model licences keep moving towards non-commercial terms or revenue-sharing requirements.
- Whether small models retain their download dominance as local hardware improves.
- Whether likes and downloads begin to converge for newer releases.
Discussion spark: When judging an open model, which signal should carry more weight for you: downloads, derivatives, licence terms or benchmark results?
Sources and evidence
- State of Open Models: Summer 2026 Observations (3 September 2026, 09:25 UTC)
not affiliated with or endorsed by Hugging Face