Hugging Face’s summer 2026 open-model report separates two signals that often get mixed together: the models people pay attention to and the repositories production systems actually use.
The report covers observations from January through August 2026. Hugging Face says public model repositories grew from 2.43 million to 2.96 million over the period. Datasets grew from 711,000 to 1 million, and Spaces grew from 1.00 million to 1.44 million.
Under that growth, usage remains highly concentrated. Hugging Face says roughly 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.
That is the most useful number in the report. The open-model ecosystem is expanding, but the production dependency layer is much narrower than the release count suggests.
Likes and downloads tell different stories
Hugging Face compared the top 25 model repositories by 2026 downloads with the top 25 by likes. It says exactly one repository appeared in both lists.
The split is intuitive once stated. A like records attention around a release. A download records that a model is wired into a pipeline, dependency, demo, scheduled job, evaluation run, or application. New frontier-scale models can dominate discussion while older small models continue to move most of the infrastructure traffic.
Hugging Face says no model published in 2026 reached the top 25 by downloads in the period it studied, while 13 of the top 25 date from 2022. It cites all-MiniLM-L6-v2 as receiving 1.55 billion pulls in seven months while collecting 5,156 likes.
For coverage, rankings, and procurement, that matters. A release can be strategically important without being widely deployed. A boring embedding model can be more operationally important than a flashy frontier release because it sits inside thousands of pipelines.
Open weights are splitting into strategies
The report also draws a sharp distinction between large frontier-focused releases and full-spectrum model families. Hugging Face says Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70 billion parameters, while Tencent and Alibaba Qwen cover a wider range from small models upward.
That is not only a technical difference. It is a distribution strategy. A frontier-only open release competes for benchmark position, developer attention, and API demand. A family that spans small, medium, and large models competes to become the default choice across local devices, enterprise deployments, and hosted inference.
Hugging Face says Qwen’s broader release strategy reached about 2,045 million downloads across repositories with declared parameter counts, compared with 37 million for Moonshot’s frontier-only portfolio. That is not a quality verdict. It shows how model size coverage changes adoption.
The report also notes licensing tension. It says 59% of Chinese releases above 20 billion parameters in 2026 used Apache 2.0 and 22% used MIT, while recent very large models such as Kimi K3 and Qwen 3.8 2.4T started to add non-commercial restrictions or revenue-share requirements.
