How Open Source AI Models is built
Measured numbers
Counted by the collection, not written by hand. A figure we cannot support is left out rather than shown as zero.
- Pages
- 148
- 12th of 18 measured in Tech/AI, by sitemap URLs
- Page count over time
- Page count has not changed: 148 pages measured, 30 Sep to 25 Oct
- Languages
- English
- Deploys
- 5 in the last 30 days
- Last deploy 22 Oct 2026. From the public status feed
- llms.txt
- Published
- llms.txt 4.5 KB · llms-full.txt 76.5 KB
- Feed
- Published
- Sitemap
- Listed in robots.txt
- Content last updated
- 14 Jul 2025
- As reported by the site itself
Tech and AI search
Read from the live site on 2026-10-23: its homepage, robots.txt and llms.txt. A line we could not read is left out.
- Built with
- Next.js
- Read from the homepage markup
- Structured data
- Organization, WebSite
- schema.org types on the homepage
- AI crawler access
- Allows all 11 major AI crawlers
- From the site’s robots.txt
- llms.txt link depth
- 30 page links
- Markdown links in the site’s llms.txt
- Search
- Search box on the homepage
How it is built, by role
The same roles apply across the whole collection. Nothing in this block is specific to this site; what the site says about itself is quoted word for word below it.
- 1
Edge network
Visitors reach a site through an edge layer: a content delivery network with a web application firewall in front of it.
- 2
Application host
The edge talks to an application host that serves the site and renders its pages.
- 3
Data
What the pages show comes from a relational database or, for a site pre-rendered at build time, from data files built into its release.
- 4
Scheduled measurement
The collection measures its sites on a schedule: content dates and site health nightly, page counts read from each site’s sitemap weekly. Every reading is kept with its date. A site too large to re-walk is not re-measured, and shows no trend.
- 5
Public status feed
A monitoring panel sends a signed report on uptime, deploys and backups once a day, and the Operations page publishes it. A figure on this page that comes from that report says so.
In the site's own words
Each passage is copied unaltered from a page the site publishes, with the address it was taken from and the date we took it. We add nothing between them.
“Open Source AI Models is an independent, curated directory of open-source large language models.”
“We track parameters, context length, licence, benchmarks (MMLU, HumanEval, MATH and others), quantization options, VRAM requirements, hardware recommendations, and inference pricing across the major hosted providers.”
“We don't cover closed-API-only models. If you can't download the weights, it's not in the directory.”
“We use the term broadly. Truly OSI-compliant open-source models (Apache 2.0, MIT) are listed alongside source-available models (Llama Community License, Gemma Terms of Use, Qwen License). The distinction matters for commercial use, so we flag it on every model and on each licence's detail page. See the licence comparison page for the breakdown.”
“We don't run benchmarks ourselves. All scores are from official model cards or papers.”
“We don't take payment from labs or hosting providers to rank their models higher.”
“The directory is updated as new releases ship — typically within a week of any major model announcement.”
“Where we hold a figure we haven't been able to confirm, or a provider has stopped publishing a price, we say that instead of showing a date.”
Screenshots
Timeline
Page count has not changed: 148 pages measured, 30 Sep to 25 Oct
| Date | Pages |
|---|---|
| 2026-09-30 | 148 |
| 2026-10-10 | 148 |
| 2026-10-18 | 148 |
| 2026-10-25 | 148 |
- 30 Sep 2026148 pages
- 10 Oct 2026148 pages
- 18 Oct 2026148 pages
- 25 Oct 2026148 pages