Hugging Face Hub passes 1 million public models

Background

A trained AI model, on disk, is a bundle of files: weights, configuration, a tokenizer, and the code needed to run it. Sharing one with other people means hosting those files somewhere and publishing a page that explains what the model is and how to use it. The Hugging Face Hub became one of the main places that does this: people upload models, datasets, and small demo applications called Spaces, and anyone else can browse, download, or call them from their own code. Because uploading is open to the public, the number of entries on the Hub works as a rough, checkable indicator of how much is being shared openly.

The milestone

In September 2024 the Hub passed 1 million downloadable public models for the first time^[1]. The milestone was announced on social media on 26 September 2024 with the message “1M Models on Hub”, citing data from Hugging Face Hub Stats^[1]. Hub statistics put the count at 1,193,810 public models as of 1 September 2024^[3], so by the time of the announcement the figure had already moved somewhat past the round number.

The other numbers

Models are not the only thing on the Hub, and the other two categories moved differently.

Public datasets grew from 670 in April 2022 to more than 17,000 by September 2024^[3]. A dataset is the collection of text, images, audio, or other material used to train or evaluate a model; the growth suggests that what gets shared openly is not only the finished models but also the raw material behind them.

Spaces are demo applications that run in a browser tab. They stood at 17,291 in September 2024, after peaking at 22,227 in August 2023^[3]. Unlike models and datasets, that figure was below its earlier high. Read together, models and datasets were climbing while the number of demo apps had risen and then come back down.

Why it is worth recording

First, it turns “how many public models are there” into a number that can be compared over time. A specific figure is easier to remember, and easier to check later, than a general impression.

Second, it reflects a change in how models circulate: not only through papers and a handful of download links, but as packages that are uploaded, revised, and replaced continuously. When model files are published this way, comparing two models becomes less a matter of reading a description and more a matter of downloading both and trying them.

Third, the number has limits. It counts entries uploaded to the Hub, not how often each one is used, and not how good it is; different versions and different fine-tuned variants of the same model each occupy their own entry. One million describes the scale of public sharing, not the strength of any particular model, and a large catalogue also makes the problem of finding the right model more pressing rather than less.

After

Growth did not stop at a million. By October 2025 the Hub held over 2 million public models^[2], roughly a doubling in about a year. The million-model mark is therefore best read as a point on a curve rather than an endpoint: it records how large the public catalogue had become in September 2024, at a moment when model files had become ordinary things to publish and download.