The most used artificial intelligence model in the world does not write, draw or talk. It is all-MiniLM-L6-v2, a small model that turns a sentence into a list of numbers so software can tell whether two texts mean the same thing, and in the last thirty days it was downloaded 251,048,129 times from Hugging Face. That is 2.83 times the second most downloaded model on the platform. We read five public rankings from the Hugging Face Hub on 21 September 2026 and they tell one story: what the industry actually runs and what the industry actually admires are two different lists.
The busiest models are plumbing
Among the fifteen most downloaded models, eight of the fifteen do the same unglamorous job: they convert text into vectors for search, ranking or classification. Only one of the fifteen, Qwen3-0.6B, is the kind of model that writes an answer to a question. Google's bert-base-uncased, first published years ago, is still fifth on 46,140,403 downloads a month and has served 3,238,101,939 in its lifetime. Together the fifteen drew 758,585,348 downloads in thirty days, and the leader alone accounts for 33.1 per cent of them.
A download here is an HTTP request to a repository's query file, counted by Hugging Face on its own servers. That is why the number is enormous: an automated pipeline that pulls an embedding model on every build is counted on every build.
The admired models are the famous ones
Sort the same platform by likes and the list changes completely. The most liked models are led by Qwen3.8-27B on 15,900 likes, followed by the image model FLUX.1-dev and DeepSeek-R1. Five of those fifteen generate images or video. all-MiniLM-L6-v2, the busiest model on the platform, sits eleventh on this list with 6,095 likes. Downloads measure use; likes measure attention, and the two rankings share almost nothing.
Everything released in 2026 has a Qwen version number
Screening the 5,000 most downloaded repositories on the Hub for the ones created this year leaves 1,385 candidates, and the fifteen busiest 2026 releases are dominated by one family: nine of the fifteen carry a Qwen version number, seven of them published by Qwen itself and the rest repackaged by NVIDIA and unsloth. The leader is not even a lab's own upload. It is Comfy-Org's single-file build of MiniMax-H3 on 20,700,265 downloads, packaged for ComfyUI, while MiniMaxAI's own repository of the same model sits further down the platform on 4,057,444.
The data everyone likes is old, and the apps everyone likes are leaderboards
The most liked datasets are headed by something that is not training data at all: prompts.chat, a single CSV file of prompt texts, on 9,839 likes, 2.93 times the web-scale corpus fineweb. Twelve of those fifteen repositories were created before 2024 and none of them in the last year. On the application side, the most liked Spaces are led by deepsite on 16,610 likes, three of the fifteen are leaderboards rather than tools, and eight are built with Gradio. Thirteen of those fifteen were actually running when we read the list; one was paused and one reported a configuration error, and both keep their likes.
| Ranking | Leader | Figure |
|---|---|---|
| Most downloaded models | all-MiniLM-L6-v2 (sentence-transformers) | 251,048,129 downloads in 30 days |
| Most downloaded 2026 releases | MiniMax-H3 (Comfy-Org) | 20,700,265 downloads in 30 days |
| Most liked models | Qwen3.8-27B (Qwen) | 15,900 likes |
| Most liked datasets | prompts.chat (fka) | 9,839 likes |
| Most liked apps (Spaces) | deepsite (enzostvs) | 16,610 likes |
None of these numbers is a measure of quality, and none of them counts people. A download is a request, a like is an account, and both are published by the platform itself without any filtering for automation. What they do show is the shape of the open model ecosystem in September 2026: a long tail of famous models that everybody applauds, sitting on top of a handful of small, old, boring models that everybody actually runs.