Hugging Face Finds Open-Model Attention and Adoption Barely Overlap
A January-to-August Hub analysis separates launch excitement from the smaller, older models that remain embedded in real developer workflows.
Hugging Face published a January-to-August 2026 analysis on August 14 that draws a sharp line between attention around open-model launches and the models developers appear to use repeatedly. Only one repository appeared in both its top 25 by downloads accumulated during 2026 and its top 25 by likes, suggesting that excitement and operational use are measuring different things on the Hub.
The report matters to developers, model publishers and buyers because leaderboards and launch-day engagement can make the newest frontier release look like the centre of the ecosystem. Hugging Face's own data instead shows a layered market: large new models attract attention, while smaller and older models remain wired into production pipelines. The figures are platform observations, not a complete measurement of commercial adoption or model quality.
The Hub grew, but usage stayed highly concentrated
Hugging Face says public model repositories increased from 2.43 million to 2.96 million during the period. Public datasets rose from 711,000 to 1 million, while Spaces grew from 1.00 million to 1.44 million. The supply of public artefacts is expanding quickly, but that does not mean use is distributed evenly across them.
The report says roughly 85.6% of models have fewer than 200 lifetime downloads and 1.5% of repositories account for 99.2% of downloads. Those platform figures describe a steep concentration of observed activity. They also explain why a stream of new releases can coexist with a relatively stable set of models carrying most routine workloads.
Likes and downloads answer different questions
A like is an attention signal: it tends to arrive when a release is new and visible. A download can recur whenever a scheduled workflow, application or build retrieves a model. Hugging Face found that no model published in 2026 reached its download top 25 for the period, while 13 of those 25 repositories dated from 2022.
That does not prove the older models are better. It shows that replacing embedded infrastructure takes longer than attracting interest. The practical implication is that teams comparing open-model ecosystems should separate launch momentum from installed use, and should avoid presenting either likes or downloads as a universal adoption score.
Small models remain the working layer
Among repositories that declare parameter counts, models below 1 billion parameters accounted for 83% of all-time downloads in Hugging Face's analysis. Models above 100 billion accounted for 1%. Restricting the view to download volume accumulated during 2026 changed the shape only slightly: 3% went to models above 70 billion parameters.
The report connects that pattern to deployability. Smaller models fit more hardware and more constrained applications. Large open-weight releases can still matter for frontier research, hosted APIs and community quantisation, but their visibility should not be confused with the volume of local and routine use recorded by the Hub.
Derivatives show where developers build
Hugging Face counted 151,448 Qwen-based derivatives, 2.6 times Meta's total footprint and 4.7 times the number of Llama repositories. The report attributes that position to a broad range of sizes, regular releases and permissive options that let developers remain in one model family across different workloads. The derivative count measures downstream repository activity, not revenue or unique production deployments.
This is analysis rather than a market-share declaration. Hugging Face explicitly says its downloads, likes, derivatives and release counts measure different aspects of activity. They do not capture private deployments, off-Hub distribution or API use, and they should not be treated as direct measures of quality or commercial adoption. The next useful checkpoint will be whether the same attention-versus-use gap persists in the next Hub report and whether off-platform evidence points in the same direction.
Status
Analysis. Internal confidence is medium because the measurements and interpretation come from Hugging Face's own platform analysis without an independently fetched audit. Questionable licensing claims elsewhere in the source were excluded from this article.
Sources
Update note: Last reviewed 2026-08-16. We will revise this analysis if Hugging Face corrects the dataset, publishes a reproducible methodology package or releases a later ecosystem snapshot.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.