Hugging Face Hub · 27 vision pipeline tags in 8 fields
The State of Open Source Vision
103,703open vision models, plus 124K in generation
49Kadded in the last 12 months
1.7×the 12 months before that
8fields, from 124K models to 1K
74%of 30-day downloads go to transformers repos
as of 2026-10-01 · figures above leave out generation
(124K repos, 0.56× in 12 months), the
charts below include it
What open vision is made of
Every public model repo carrying one of the 27 vision tags. Pick a
field here or in the bar above to follow it through the rest of the page. Fine view splits
the cards, library bars and table into single tags.
New models per month, by field
Stacked, so the column height is all of open vision and each band is one
field. Generation sets the scale; the fields moving fastest are underneath it.
Fine view: pick a field above to split this chart into its tags.
A shared axis would flatten most of these into nothing: Generation is
121× Depth & pose.
Each sparkline is scaled to its own field, so what you compare is shape, not size.
Growth is the last 12 months against the 12 before.
Generation
Vision-language
Classification
Detection
Embeddings
Segmentation
Video classification
Depth & pose
Where transformers is the library
The library each repo declares, as a share of the repos created in the last
12 months and of the last 30 days of downloads. transformers is the dark segment; the two
biggest other libraries follow, then the rest. Repo counts show who ships, downloads show
what gets used.
repos, last 12 monthsdownloads, last 30 daystf repostf dl
A contribution grid weighted by engagement, not output: labs ranked by
the likes their vision models have earned, months across, darker where a month's releases
landed. Counting repos instead would rank format converters and checkpoint farms above
every research lab — Qwen earns 65K likes from 44 models while timm earns 1.6K from 1,785. Sibling repos are collapsed into one model, so a
family of sizes or fine-tunes counts once rather than once per repo. The dot is the field
each lab mostly works in; hover a row for its most-liked model.