Which Open AI Models People Download: September 2026 Data

Ask which open AI models people actually use and most answers are opinions. Download counts are not a perfect answer, but they are a measured one. On September 18, 2026 we queried the Hugging Face Hub, the main public host for open model weights, for the most-downloaded models in its three language and vision-language categories and kept the top 60 by downloads over the previous 30 days. Qwen repositories hold 27 of the 60 rows. Qwen3-0.6B, a model small enough to run on a phone, leads at 22.5 million.

This page is for anyone choosing an open model who wants to know what the rest of the ecosystem is pulling, and for writers who need a dated, reproducible adoption table rather than a vibe. It explains what a Hugging Face download does and does not count, prints all 60 rows, sums them by model family, and says plainly where downloads stop being evidence. The query is stated in the methodology so anyone can repeat it.

The Hugging Face Hub reports a "downloads" number per model repository. In the Hub API this figure is a rolling count over the last 30 days, which is why it moves every day and why the date of collection matters. It counts file fetches from the repository, which means a person downloading a model once, a server pulling weights on every fresh deployment, and a continuous-integration job fetching a tiny model to run a test all add to the same number. It does not count a model that was downloaded once and then copied around a company, and it does not count use through an API. The Hub's API documentation describes the endpoint and its sort options.

Three consequences follow, and they shape how to read the table. Small models rank high because they fit in more places and are cheap to re-download. Test fixtures rank high because CI runs many times a day. And a quantised repack of a popular model counts separately from the original, so a model family's real reach is the sum of its rows, not its best one. We flag the fixtures with a dagger, keep every row otherwise, and sum by family in section 03.

Rank is by rolling 30-day downloads at collection time. Model is the repository name, organisation included, exactly as the Hub returns it. Licence is the tag the repository declares; "Other (custom)" is the Hub's own category for a licence file that is not a standard one. A dagger marks a repository we consider a test fixture: a model whose downloads come mainly from automated test suites.

Summing rows by the model they derive from, and leaving out the three fixtures, gives the chart below. We assigned a family by the repository name and its declared base model, so a community GGUF of Qwen3.8-27B counts as Qwen and Nvidia's NVFP4 build of Qwen3.6-35B-A3B counts as Qwen. The 10 largest families are shown; the remaining three rows, Kimi K3, Florence-2 and Chandra OCR, are each a single repository under 3.4 million.

Two families deserve a note. Ornith's three GGUF repositories sum to 13.5 million, which puts a name most readers will not know above Llama and gpt-oss; we have not investigated the source of those downloads and record the number as the Hub reports it. And Llama has two rows in the top 60, both from the 3.x generation released in 2024, which fits the shift toward Qwen and DeepSeek described in our first-half 2026 open-weight retrospective.

The repack finding is the practical one. If you want to know which size and format of a model people actually run, the quantised rows answer it more directly than the originals: an FP8 or 4-bit file is what gets loaded onto a real machine. Our self-hosting guide for open coding models pairs those formats with hardware, and this week's comparison of SSD streaming and ternary weights covers two newer ways to make a large model fit; the Edge0 model from that post was the Hub's top trending repository on the day we collected, with 37,131 downloads and 3,349 likes.

Downloads are not users, and users are not production. A download count says how many times files were fetched from one host in 30 days. It cannot distinguish a person from a script, a trial from a deployment, or ten thousand developers from one company's build farm. It excludes every model served through an API, every mirror such as ModelScope, and every weight file that was fetched once and then cached inside an organisation. A model that is downloaded by a few hundred companies and run at scale can rank below a model that is fetched by a hundred thousand hobbyists once each.

It also cannot tell you where use is. We have not computed a country share and do not print one: a repository's organisation says where a model was made, not where it is run. Treat the table as a measure of distribution, read it alongside benchmark and price data, and check the licence column before assuming any row is usable in a product. If you want help choosing an open model for a specific workload, our AI transformation service includes that evaluation.