Hugging Face published a summer 2026 accounting of its own hub, covering January through August, and the numbers separate what developers praise from what they actually run. Hugging Face compared the 25 repositories with the most downloads this year against the 25 with the most likes. Only one name showed up on both rankings.
The gap is stark at the extremes. all-MiniLM-L6-v2, an embedding model most builders have never heard discussed, racked up 1.55 billion downloads across seven months while collecting just 5,156 likes. Kimi-K3, Moonshot’s frontier release that generated plenty of attention on launch, earned roughly 60 downloads for every like it received. Likes cluster around frontier launches during the fortnight or so that follows one. Downloads accrue to small, unglamorous models wired into a production pipeline for years. A team deciding what to standardize on should weight the second number far more than the first.
That distinction explains why Alibaba’s Qwen family, not any single frontier release, has become the platform’s real center of gravity. Developers have built 151,448 derivative repositories on top of Qwen models, a figure 2.6 times larger than Meta’s total footprint and 4.7 times Llama’s specifically, expanding at roughly 180 to 210 new repositories a day for seven straight months. Moonshot’s frontier-only catalog, by contrast, drew 37 million downloads all year. Qwen’s strategy of covering every size class drew 2,045 million, roughly 55 times as many. A steady release cadence and Apache 2.0 licensing, not any one model’s benchmark score, appear to be doing the work.
Licensing itself is shifting under that Apache 2.0 baseline. Hugging Face’s data shows Chinese labs putting frontier-scale weights under terms as open as the ones they give their tiniest releases: DeepSeek and Z.ai, for instance, license weights between 700 billion and 1.65 trillion parameters under plain MIT terms. American labs releasing models of comparable size split closer to 29 percent Apache or MIT, 41 percent under custom terms, and 30 percent with no license declared at all. But the report flags a reversal at the top of that range: Kimi K3 and Qwen3.8 have begun attaching revenue-share clauses and noncommercial-use limits to their biggest releases, a sign permissive terms were a phase, not a promise.
That reversal is worth watching against today’s other open-weights news. Z.ai’s GLM-5.3, an open model claiming state of the art coding performance with weights promised in two weeks, and Alibaba’s Qwen 3.8 27B, released under Apache 2.0, both extend the pattern this report describes. Frontier-scale open releases increasingly originate from Chinese labs, and the terms attached to the largest of those releases are no longer guaranteed to stay permissive.
Hugging Face is grading an ecosystem it also hosts, which makes its download and derivative counts the most complete data available anywhere on model adoption. It also means the view stops at the Hub’s edge. API traffic, private deployments, and models distributed through other channels do not show up in these numbers, a limitation the report itself acknowledges.
For teams choosing a base model to standardize fine-tuning and deployment around, the report’s actionable signal is to check derivative count and download trend rather than launch-week buzz. For anyone assuming an open Chinese release will stay permissively licensed a year out, the Kimi K3 and Qwen3.8 changes say to verify terms release by release, not by lab.
Hugging Face published its State of Open Models: Summer 2026 report on August 14, 2026.
