Hugging Face Hub

What it is The Hugging Face Hub is the platform that Hugging Face runs for hosting and sharing machine-learning models, datasets and demo apps, which are called Spaces. The company is often described as the GitHub of machine learning, a comparison to the code-hosting service that developers use to share software.^[1] Machine-learning models are files of learned numerical weights, and a dataset is a collection of examples used to train or test a model. A repository on the Hub holds the files for one model, dataset or Space, with its history and documentation. ## Origins The Hub's repositories were built on Git, the version-control system, together with Git LFS (Large File Storage), an extension for large binary files. In 2020 the company released the huggingface_hub Python client library, with its first release (v0.0.1) arriving in late 2020.^[1][2] The client is the code that lets a program or a person upload and download files from the Hub. Gradio, a Python library for building demos of machine-learning models, joined the company in December 2021. Its founder said more than 300,000 demos had been built with it by then.^[4] Gradio became the basis for building demo apps on the Hub. In June 2022 the client library gained an HTTP commit API. Uploads therefore no longer needed Git or Git LFS.^[2] This removed a technical barrier for people who wanted to publish large files without learning the Git workflow. ## Growth The Hub passed 1 million public models in September 2024, and the milestone was announced as “1M Models on Hub” on 26 September 2024.^[3] In August 2024 the Hub held about 1.3 million models, 450,000 datasets and 680,000 Spaces. Its large-file storage held about 12 PB, or petabytes, across 280 million files, and it served about 1 billion requests a day.^[1] A petabyte is a thousand terabytes, which indicates how much data the platform was storing at that point. ## Storage: from Git LFS to Xet Hugging Face acquired XetHub in August 2024 to replace Git LFS with chunk-level, deduplicated storage. In practice this means that when a large file is edited, only the changed chunks need to be uploaded again, rather than the whole file.^[1][2] Deduplication means that identical pieces of data are stored once. By late 2025 the Xet migration covered over 77 PB across more than 6 million repositories.^[1][2] ## Scale in late 2025 By October 2025 the Hub held over 2 million public models, 500,000 public datasets and 1 million public Spaces. Including private repositories, the totals were roughly double.^[2] The huggingface_hub v1.0 release shipped on 27 October 2025 with a new hf command-line tool.^[2] A command-line tool is a program that is run from a terminal rather than through a graphical window. ## Features added over time Several services have been added to the Hub over time. They include Inference Endpoints, which host a model for use in an application; Inference Providers, which offer one API across several serverless providers; a Jobs API introduced in the third quarter of 2025; and integration with MCP, the Model Context Protocol that lets AI assistants connect to outside tools and data.^[2] Serverless means that the user does not manage the servers that run the model. ## Growth to 2026 By September 2026 the platform had over 3 million models and more than 18 million users. These figures were cited when Nvidia agreed to acquire Hugging Face.^[5] ## Security incident in July 2026 In July 2026 Hugging Face disclosed a breach of its production systems. The breach began with a malicious dataset in its processing pipeline. The company said that public models, datasets and Spaces showed no sign of tampering, and it advised users to rotate their access tokens.^[6] An access token is a secret string that grants a program or a person permission to use an account, so rotating it means replacing it with a new one so that the old one no longer works. ## Impact The Hub is one of the main places where open machine-learning models are published and downloaded. Its storage changes, client libraries and security practices affect a large number of developers, which is why its growth figures and incidents are recorded here in detail. Readers who need exact numbers for a particular date should check the figure against the date it was reported, because the totals rose quickly over these years.