Where open weight models actually stand
Two facts about who is publishing reframe the year more than any benchmark delta does.
Chinese labs hold the size ceiling
In almost every month of 2026, the largest and most capable open model from a Chinese lab was bigger than anything an American lab released of its own weights. This is a statement about the ceiling, not the average.
The distinction matters when reading the open-versus-closed comparisons that dominate English coverage. Those comparisons ask how far behind the closed frontier the best open model sits. The Hub data answers a prior question: whose weights are setting that open ceiling in the first place — and for most of 2026 the answer has not been an American lab.
"In almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than anything an American lab released of its own."
The most prolific publishers make chips
The two organizations publishing the most new open models this year are AMD and NVIDIA. Neither is a model lab in the sense the phrase usually carries.
It follows a commercial logic that is easy to miss. Hardware vendors need models tuned and packaged for their own accelerators, because a model that runs well on your silicon is a reason to buy the silicon. The result is that a large share of new repositories are optimization artifacts rather than new capabilities — which is another reason repository counts do not measure progress.
"The two organizations publishing the most new open models this year are also the companies making the hardware: AMD and NVIDIA."
What downloads and likes actually measure
Both numbers are printed next to every repository, and they are routinely quoted as popularity. They are not measuring the same population at all.
The two leaderboards share one entry
Of the top 25 repositories by downloads this year and the top 25 by likes, exactly one appears on both lists. A metric pair that disagrees this completely is not two views of popularity.
Downloads accrue to whatever an automated pipeline pulls on a schedule: an embedding model called on every document, a tokenizer, a classifier inside a service. Likes are cast by people reacting to a release. One measures machine consumption, the other human attention, and quoting either as adoption imports the wrong one.
"We took the top 25 model repositories by downloads accumulated this year and the top 25 by likes. Exactly one repository appears in both lists."
Almost all volume goes to models under 1B
Models under 1B parameters take 83% of all-time downloads, while everything above 100B takes 1%. The models that headline every capability discussion are a rounding error in the traffic.
This is the clearest evidence that download counts are a deployment signal rather than a quality one. Small models are what fits inside a product: they run on a CPU, cost nothing per call, and get pulled on every build. Anyone using download rank to argue that one frontier model beat another is reading a chart about embedded infrastructure.
"Among models that declare a parameter count, those under 1B take 83% of all-time downloads and everything above 100B takes 1%."
The report disclaims the reading most people give it
Hugging Face states plainly that these metrics are not direct measures of quality, commercial adoption or market share. The caveat is worth repeating because the numbers are so quotable without it.
What the data does describe is activity on one platform: what gets published there, pulled from there and built on there. That is a real and useful population — it is where open model work happens — but it excludes closed APIs entirely and says nothing about enterprise spend.
"They should not be interpreted as direct measures of model quality, commercial adoption, or overall market share."
Qwen became the thing everyone builds on
The most consequential number in the report is not about capability. It is about what people fine-tune when they need a starting point.
151,448 derivatives, and the gap is not close
Qwen-based models account for 151,448 derivatives on the Hub — 2.6 times Meta's entire footprint and 4.7 times the Llama repositories specifically. Google follows at a distance.
A base model position compounds in a way benchmark scores do not. Tooling, quantized builds, fine-tuning recipes and community knowledge accumulate around whatever people already use, and each of those makes the next team's choice easier. Llama held this position; the Hub data shows it changed hands during 2026.
"Qwen-based models now account for 151,448 derivatives on the Hub, 2.6× Meta's total footprint and 4.7× the Llama repositories specifically."
Agent traffic is now measurable, and partly unaccounted for
The Hub began recording the identifying token that coding agents send when they call it, which makes agent traffic visible for the first time. The early reading is that the field is wider than the named list.
Nearly a quarter of agent-tagged traffic in July came from harnesses the dataset does not yet name. For anyone tracking which coding agents matter, that is the more informative number: the tools in the published breakdown are not the whole market, and a quarter of the traffic belongs to something else.
"Nearly a quarter of agent-tagged traffic in July came from harnesses not yet named in the dataset."
The report itself is a long page of tables and stacked charts, and the caveats live in the prose between them rather than in the figures. Converting it to markdown before handing it to a model keeps each number attached to the sentence that qualifies it, which is precisely what gets lost when a chart is screenshotted into a slide.
If you are deciding what to build on, the Hub data is more useful than the gap charts. It says the community has consolidated on one base model family, that the largest open weights are coming from Chinese labs, and that the models carrying real production traffic are small enough that no benchmark article mentions them. None of that shows up in a scoreboard measuring the distance to the closed frontier.
