459 repos hold half the stars in open-source AI
The catalogue holds 10,840 projects and 58,331,228 stars. Cross the list from the top and you reach half of all those stars at project number 459 — 4.2 % of the shelf. The top 1 % hold a quarter; the bottom half hold 5 %.
TL;DR — the short version
459 repositories hold half the stars. The 10,840 open-source AI projects in the olud.ai catalogue carry 58,331,228 stars between them, and 4.2 % of the projects account for 50 % of that total.
The top 1 % — 108 projects — hold 25.0 %. The top 5 % hold 53.7 %, the top 10 % hold 69.0 %.
The bottom half holds 5.0 %. 5,420 projects share 2,943,548 stars; 5,096 projects sit below 1,000 stars.
What follows for you: stars can confirm a shortlist, but they cannot build one. Below the top few percent the differences are noise — rank on fit, licence and pulse instead.
Every catalogue of software eventually gets asked the same question: what is popular? We can answer it precisely, because the olud.ai catalogue now holds 10,840 open-source AI projects with 58,331,228 stars between them. The answer is less useful than it sounds, and the reason is the shape of the distribution rather than the size of it.
A mean four times the median
The average project in the catalogue has 5,381 stars. The median project has 1,134. When the mean sits 4.7× above the median, you are not looking at a spread of similar things — you are looking at a handful of giants and a very long shelf of everything else. Sorting by cumulative share makes the gap explicit.
Cross the list from the top and you reach half of all the stars in the catalogue at project number 459. Four percent of the shelf, half the applause.
No single repo owns the top
Concentration at this scale usually comes with a monopoly at the very top. Here it does not. The largest project in the catalogue holds less than one percent of all stars.
| Project | Stars | Share of the catalogue |
|---|---|---|
| sindresorhus/awesome | 501,710 | 0.86 % |
| public-apis/public-apis | 473,758 | 0.81 % |
| openclaw/openclaw | 383,287 | 0.66 % |
| nilbuild/developer-roadmap | 365,995 | 0.63 % |
| vinta/awesome-python | 317,393 | 0.54 % |
Three of those five are curated lists rather than software, which is its own comment on what a star measures: the cheapest thing to star is a page of links. The concentration is therefore broad rather than monarchic — it takes a hundred projects to make a quarter of the total, and no one of them dominates.
It is also worth noting how the top is composed. Ownership is barely concentrated at all: the catalogue spans 8,109 distinct owners, 6,861 of whom publish exactly one project, and the ten largest owners together account for only 8.8 % of all stars. Stars pile up on projects, not on publishers.
The bottom half is not empty — it is unread
5,096 projects in the catalogue sit below 1,000 stars, and 170 below 100. It is tempting to read that as a quality signal, and mostly it is not one. A tool published this spring for a narrow task has had neither the time nor the audience to collect stars, and its 300 stars say nothing about whether it works. The same 300 stars on a five-year-old repository say something quite different.
This is the practical failure of star-based ranking: the number conflates age, audience and merit into one figure and then hides which of the three is doing the work. Two projects with 1,100 stars each — both sitting on the median — can be a finished utility with a stable user base and an abandoned experiment that trended for one week in 2023.
Why the shape matters beyond ranking
A distribution this steep also explains why open-source AI feels like it has fewer options than it does. Discovery mechanisms — trending pages, conference talks, blog posts, model cards — all sample from the top. If 69 % of the visible mass belongs to a thousand projects, then the other 9,756 are effectively invisible even when they fit a problem better. That is the gap a catalogue is supposed to close, and why we publish the whole shelf rather than a top ten. If the project you find at the top has gone quiet — and 42 % of the catalogue has — the replacement almost always exists further down.
Method
Figures cover the 10,840 projects in the olud.ai catalogue, snapshot of 24 September 2026, with 58,331,228 stars in total. Slices are cut by star rank: the top 1 % is the 108 most-starred projects, the top 5 % the 542 most-starred, and so on; shares are the sum of stars in the slice divided by the catalogue total. The median is the star count of the project at rank 5,420. Star counts are as reported by each project’s forge and are never added across forges — a GitHub mirror of a GitLab project is counted once, at the forge we track it on. Curated lists are included in every count, as they are part of what people actually star.
Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.
Keep reading
- One fork for every 6.9 stars — the second number on every repo page, and what it tells you that the star count cannot.
- 127 repos above 10,000 stars are held up by fewer than ten people — how thin the maintenance layer really is.
- 42 % of open-source AI has gone quiet — the census of last-push dates across the same catalogue.