Hugging Face Opened Its Hub Data for January to August 2026: 1.5% of Repositories Took 99.2% of All Downloads and 83% of Downloads Went to Models Under 1B Parameters

Hugging Face published a mid-year state of open models report on August 14, 2026, covering seven months of Hub activity from January to August 2026, and reported that 1.5% of repositories accounted for 99.2% of all downloads while 85.6% of repositories never reached 200 lifetime downloads. Over the same period 83% of downloads went to models under 1 billion parameters, and models above 100 billion parameters took 1%. Public model repositories grew from 2.43 million to 2.96 million, and datasets grew from 711,000 to 1 million. ASAP works only from figures verifiable in that report, then examines what the same numbers do not say about the open model ecosystem.

The monthly ceiling for Chinese releases ran between 754 billion and 2.78 trillion parameters

Hugging Face compared the largest model published each month by country and reported that China's monthly ceiling ran between 754 billion and 2.78 trillion parameters while America's ceiling stayed under 130 billion in five of the seven months. The large Chinese releases named in the report are Moonshot's Kimi-K3 at 2.8 trillion parameters, Alibaba Qwen's Qwen 3.8 Max at 2.4 trillion, and DeepSeek-V4-Flash at 284 billion.

Large American open releases exist in the same table. Thinking Machines Lab's Inkling is 952 billion parameters, NVIDIA's Nemotron 3 Ultra is 561 billion and its Nemotron 3 Super is 124 billion, and Arcee AI's Trinity-Large is 399 billion. What the monthly ceiling comparison exposes is not the absence of large American models but the absence of a monthly cadence of them.

Size is not quality, and the report says so directly. This tally counts parameters of open weights uploaded to the Hub and is not a measure of model quality or commercial adoption. The contrast therefore indicates which side pushed large weights into the open more consistently, not which side built better models.

1.5% of repositories took 99.2% of downloads

Hub downloads in 2026 are concentrated to an extreme degree, with 1.5% of all repositories accounting for 99.2% of downloads while 85.6% of repositories never passed 200 lifetime downloads. By size band, models under 1 billion parameters took 83% of all-time downloads and models above 100 billion took 1%, while within 2026 alone models above 70 billion accounted for 3% of volume.

Attention and adoption diverge in the same data. Exactly one repository appeared in both the top-25 downloads list and the top-25 likes list over the seven months. The sentence embedding model all-MiniLM-L6-v2 was pulled 1.55 billion times against 5,156 likes, while Kimi-K3 drew roughly 60 downloads per like.

The Hub itself kept expanding. Public models rose from 2.43 million to 2.96 million, about 21.8%, datasets from 711,000 to 1 million, about 40.6%, and Spaces from 1.00 million to 1.44 million, or 44%. Placing 21.8% repository growth next to a download share of 1.5% shows that most newly created repositories added entries to a list rather than users to a model.

Qwen took the community base model position

In the derivative tally, Alibaba's Qwen family holds 151,448 derivative repositories on the Hub, 2.6 times Meta's total footprint and 4.7 times the Llama repositories specifically. New Qwen-based repositories appear at roughly 180 to 210 per day, and GGUF format downloads run at 39.6 million per month.

That position used to belong to Llama. When the base model that accumulates fine-tunes, quantizations, and adapters changes, the tooling, recipes, and tutorials built on top migrate with it, which makes derivative counts a measure of ecosystem inertia rather than popularity. Hugging Face attributes Qwen's position to a broad model family, a consistent release cadence, and openness, not to any single model's benchmark score.

Licensing split in the band above 20 billion parameters

Of 178 Chinese releases above 20 billion parameters this year, 59% carry Apache 2.0 and 22% carry MIT, and none carry non-commercial restrictions. In the same size band, American releases carry Apache or MIT 29% of the time, custom terms 41% of the time, and no specified license 30% of the time.

The 30% unspecified figure matters more in practice than the headline percentages. Weights without a stated license rarely clear internal review, and the 41% under custom terms require reading usage scope and redistribution conditions case by case. In organizations where adoption speed is set by the legal review queue rather than by benchmarks, this table reorders the candidate list directly.

Claude Code held 44.4% of July agent traffic

Hugging Face broke out agent traffic for the first time in this report and found that Claude Code held 44.4% of it in July, Codex held 20.8%, and nearly 24% arrived from harnesses not yet registered by name. Claude Code's share stood at 67.8% in April, a drop of 23.4 percentage points in three months, and unregistered traffic peaked at 59.8% in May.

The report's framing is that agents are the new user. Alongside the path where a person reads a model card and downloads weights, there is now a path where a coding agent pulls weights and datasets mid-task. The unregistered 24% is also the blind spot of this tally: with a quarter of traffic unattributed, the two named shares are safer to read as a ranking than as exact proportions.

Small models did not win; distribution paths diverged

Reading 83% of downloads under 1 billion parameters as a victory for small models is half right at best. Runtime growth tells the other half: GGUF repositories grew 464% over the seven months, the robotics library lerobot grew 194%, and Apple's mlx grew 148%, while transformers and peft grew 16% and diffusers grew 21%. Everything that grew fast is a format or library used for local execution and on-device inference.

The scarcity of large model downloads follows from the same split. Hugging Face states that this tally excludes API usage, private deployments, and models distributed through other channels. Most organizations running models above 100 billion parameters call an API rather than downloading weights, so the 1% share here indicates where consumption happens, not an absence of demand. What the report measures well is therefore not the model market but the composition of the layer that downloads weights and runs them directly.

What a team can actually take from this table

For a team choosing a fine-tuning base, 151,448 derivative repositories is not a performance number but an inventory count of documentation, examples, and quantized artifacts. At comparable quality, the family with the thicker derivative tail shortens debugging time, while a thin family adds deployment work even when it leads on benchmarks. Qwen's intake of 180 to 210 new repositories per day indicates that gap widening rather than closing.

For teams evaluating on-device or local inference, 464% GGUF growth and 148% mlx growth are the more actionable figures. An organization running Mac-based development alongside on-premises deployment will find that support for those two formats filters the candidate list before any benchmark does. Layer the 30% unspecified-license share on top, and candidate selection becomes a license check and a format check first, with benchmark comparison after.

What this table does not measure

Hugging Face closes the report by stating that these metrics should not be interpreted as direct measures of model quality, commercial adoption, or overall market share. Downloads reflect Hub ecosystem activity, open-source AI adoption happens across many channels, and Hub activity should be viewed as one perspective on ecosystem development rather than a complete measurement.

Dropping that caveat produces predictable misreadings. Download counts rise with automated retries, cache misses, and agent calls, so they are not user counts, and likes measure human attention while diverging from usage, which this tally demonstrates with a single overlapping entry between the two top-25 lists. The country ceiling comparison counts only the single largest model each month, a different axis from release volume or real usage. One safe sentence survives all of it: in the first half of 2026 the layer that downloads weights directly shifted toward small and permissively licensed models, and a large share of the new arrivals in that layer were agents rather than people.

Source: ASAP analysis based on Hugging Face's official blog report State of Open Models: Summer 2026 Observations (August 14, 2026)