We measured 10,491 open-source AI projects. 42% have gone quiet.

Star counts only ever go up. Commits stop. That asymmetry is the whole story of this census: a repository can look like a landmark for years after the last person walked out of the building. So we ignored the stars for a moment and looked at one field — the date of the last push — across every one of the 10,491 projects in the olud.ai catalogue.

What the shelf actually holds

The catalogue spans 10,491 projects carrying 56,432,352 stars between them, with a median age of 1,147 days — roughly 3.1 years. In other words, the typical open-source AI project predates most of the tools people use it with today. That maturity cuts both ways: it means the shelf is deep, and it means a large part of it was built for a world that has since moved on.

Sorted by the time since their last push, the projects split like this.

Two facts sit side by side in that chart. First, the active core is real and busy: 1,898 projects were pushed in the last week alone, and just under half the catalogue — 5,121 projects, 48.8 % — moved in the last 90 days. Second, the long tail is not a tail. 4,436 projects, 42.3 % of everything we track, have gone at least six months without a push, and 3,562 of those have gone more than a year. The visible half of open-source AI is, quite literally, a museum wing.

Even the 10,000-star repos sleep

The comfortable assumption is that abandonment is a small-project disease — that the repos everyone has heard of are safe. The top of the leaderboard says otherwise.

Among the 1,150 repos that have crossed 10,000 stars, 197 — 17.1 % — have had no push in six months, and 272 — 23.7 % — none in three. Popularity clearly helps: 17.1 % is well below the catalogue-wide 42.3 %. But it does not come close to eliminating the risk. A meaningful slice of the famous tier is already quiet, and its star counts will keep growing anyway, because stars measure the past and commits measure the present.

This is the trap for anyone choosing tools by reputation. The projects most likely to appear in a blog post, a course, or a conference talk are precisely the old, heavily starred ones — the cohort where 23.7 % have not moved in a season.

578 entries are not code at all

One wrinkle worth naming: 578 entries in the catalogue are “awesome” lists and other curations — collections of links rather than software. For a list, silence is more forgivable; nothing has to compile. But a stale list is arguably worse than a stale library, because it silently forwards readers to projects that may themselves have gone dark. A curation last touched before an entire ecosystem layer appeared — the MCP server wave we mapped in a separate study is barely older than some of these gaps — is a map of a city that has since been rebuilt.

Even setting every one of those 578 entries aside, they cannot explain the silence on their own: the quiet cohort numbers 4,436. The bulk of what has stopped moving is real software.

Quiet is not the same as dead

A fair objection: some software is simply finished. A small utility that converts one file format to another can sit untouched for years and remain perfectly usable. With a median project age of 1,147 days, plenty of the catalogue is old enough to have reached that plateau honestly.

The problem is that AI infrastructure rarely gets to be finished. Model formats change, quantisation schemes change, CUDA versions and Python packaging change, provider APIs change. A wrapper library that has not moved in a year probably no longer speaks to the endpoints it wraps; an inference tool frozen for a year predates file formats that today’s local stacks treat as standard. In local AI especially, a season of silence is usually visible as friction the first time you try to install the thing.

So the right reading of 42.3 % is not “42.3 % of open-source AI is garbage”. It is: for any given project you are about to adopt, the base rate that it has already stopped moving is high enough that you should check — every time.

How to choose from this shelf

Three habits follow directly from the numbers. First, treat recency as a first-class filter: the 90-day band holds 5,121 projects, which is more than enough choice for almost any task, so there is rarely a reason to bet on a repo outside it. Second, when a famous project turns out to be quiet, look for a living replacement before writing your own — that is what our alternatives pages exist for, and the open-source option is usually still the economically rational one, as our open-vs-closed price study lays out. Third, remember that abandonment risk compounds across a stack: every quiet dependency you adopt is a migration you have already scheduled without picking the date.

None of this is an argument against open source — the 1,898 projects pushed in the past week are the counter-argument. It is an argument against choosing by fame in an ecosystem where 42.3 % of the shelf, and 17.1 % of the famous shelf, has gone still.

Method

The census covers the 10,491 projects in the olud.ai catalogue, snapshot of 19 August 2026. “Silent” means no push to the repository within the stated window, measured from the last-push timestamp reported by each project’s forge; pushes are a deliberately generous signal, counting any branch activity, so these figures understate abandonment rather than overstate it. The bands up to 365 days in the chart are cumulative — a project pushed yesterday appears in every one of them. Star totals are as reported by each forge and are not compared across forges. The 578 curation entries are included in all counts.

Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.

Keep reading

• The open vs closed LLM price gap, measured — what running open models actually saves, from the same catalogue.
• What runs on your GPU — a census of the local AI stack people actually install.
• The MCP ecosystem, mapped — the newest layer of open-source AI, counted before it settles.