127 repos above 10,000 stars are held up by fewer than ten people
A star count tells you how many people approve of a project, not how many can merge a fix into it. Across 1,190 projects above 10,000 stars, the median has 78 contributors, 54.6 % have fewer than 100, and 127 have fewer than ten.
TL;DR — the short version
127 projects above 10,000 stars have fewer than ten contributors, and 65 of those have fewer than five. Between them the 127 carry 3,213,297 stars.
More than half the famous tier is under 100 people. Of the 1,190 projects above 10,000 stars with a health score, 650 — 54.6 % — have fewer than 100 contributors; the median is 78.
A large part of the thin end is a new kind of repository. Of the 127, 53 were created since January 2025 and 32 have no primary language at all — prompt packs, skill files, curated markdown, where one author is the normal state.
What follows for you: read the contributor count next to the star count. For anything you import, fewer than ten people means your migration plan and the project’s succession plan are the same document.
A star count tells you how many people approve of a project. It does not tell you how many people can merge a fix into it. Those two numbers are wired to entirely different things, and only one of them decides what happens when the maintainer takes a job that forbids side projects.
We keep the second number for every project we score. Across the 1,190 projects above 10,000 stars with a health score, here is how the maintenance layer is actually distributed.
The famous tier is thinner than it looks
The headline number is the 127 projects — 10.7 % of the tier — that are above 10,000 stars and under ten contributors. But the more useful line is the median: 78. Half of the most visible projects in open-source AI are maintained by fewer people than a mid-sized engineering department, and 54.6 % of them by fewer than a hundred.
That is not a scandal. It is how open source has always worked, and a project with twelve committed maintainers is in better shape than one with four hundred drive-by contributors. The problem is the mismatch: the star count grows with visibility, the contributor count grows with commitment, and readers routinely use the first as a proxy for the second.
The thin end, named
| Project | Stars | Contributors | What it is |
|---|---|---|---|
| mattpocock/skills | 229,517 | 7 | Skill files |
| multica-ai/andrej-karpathy-skills | 204,889 | 7 | One markdown file |
| VoltAgent/awesome-design-md | 109,571 | 4 | Curated collection |
| karpathy/autoresearch | 94,406 | 9 | Python, research code |
| mlabonne/llm-course | 81,898 | 3 | Course |
| GokuMohandas/Made-With-ML | 49,176 | 1 | Course, notebooks |
| lutzroeder/netron | 33,381 | 1 | Software you run |
Six of those seven rows are documents. A skill file, a curated list, a course: one author is not a bus factor problem there, it is the format working as intended, and nothing in your build breaks if the author stops. Of the 127 thin projects, 53 were created since January 2025 and 32 report no primary language at all — the markdown-and-prompts wave that arrived with coding agents, and which has learned to collect six-figure star counts faster than any library ever did.
The last row is the one that matters. Netron is a model viewer people install and use daily, it has been maintained for years, and the forge reports one contributor. That is a genuinely single-person dependency wearing a 33,000-star badge — not a criticism of a tool that works, just a fact you should know before it sits inside a workflow you cannot do without.
Why we weight this at 30 points out of 100
Our health score gives the community signal 30 points of 100 — the same weight as recent activity, and more than momentum or maintenance backlog. The reason is that activity and contributors fail in different directions. A project pushed yesterday by its only author is alive today and gone the day he stops. A project with 200 contributors and no push in three months is dormant but recoverable, because the knowledge is distributed and someone can pick it up. Recency tells you about the present; the contributor count tells you about the survivable futures.
Together they separate the two ways software dies. 42 % of the catalogue has gone quiet, and the quiet projects with a broad contributor base are the ones a fork can actually rescue.
Method
Figures cover the 1,190 projects above 10,000 stars that carry a health score in the olud.ai catalogue — 5,173 projects are scored in total, out of 10,840 in the catalogue; scoring is being extended, and the unscored remainder skews towards smaller projects. Contributor counts and star counts in this article both come from the same health check, so they are internally consistent; 5,080 of the 5,173 checks were made between 20 and 22 August 2026, and each figure quoted is dated in the tables. Contributor counts are as reported by the project’s forge, which caps the list it returns for very large projects — every claim here is of the form “fewer than N”, which that cap cannot affect. Project categories in the table are ours, read from each repository. Current star counts, which move daily, are on each project’s page in the catalogue.
Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.
Keep reading
- 459 repos hold half the stars — the popularity curve that produces those badges.
- One fork for every 6.9 stars — the other counter on the page, and what it reveals.
- A quarter of the most-starred AI software has stopped receiving commits — what happens after the maintainers leave.