Hugging Face acquires XetHub to rebuild Hub storage

Background

Hugging Face runs the Hub, a hosting service for models, datasets and demo applications. At the time of the announcement the Hub held about 1.3 million models, 450,000 datasets and 680,000 Spaces^[1]. Most of that content is large binary data — model weights and dataset files that can run to many gigabytes. Git, the version control system these repositories are built on, was designed for text source code; the usual workaround is Git LFS (Large File Storage), which keeps large files outside the main repository and fetches them on demand. That works, but storing, versioning and transferring very large collections stays awkward, and before the deal the Hub ran on Git LFS^[1].

What happened

On 8 August 2024 Hugging Face announced that it had acquired XetHub, a company based in Seattle^[1]. XetHub was founded by Yucheng Low, Ajit Banerjee and Rajat Arya, who had built Apple's internal machine learning infrastructure^[1]. Low started the company in Seattle in 2021, after working at Turi, a Seattle machine learning startup^[3]. GeekWire described the deal as Hugging Face's largest acquisition to date and reported that the terms were not disclosed^[2].

How the technology works

XetHub's technology lets Git scale to terabyte-sized repositories by storing data in chunks and deduplicating it^[1]. Chunking means a file is split into smaller pieces instead of being handled as one indivisible object; deduplication means identical pieces are stored only once, no matter how many files or versions contain them. For model and dataset hosting this matters because large files are rarely written from scratch — a new version usually differs from the old one in small ways, and a storage layer that recognises repeated content can avoid keeping and moving full copies over and over. Hugging Face said it planned to move the Hub off Git LFS onto a Xet-based backend^[1].

What changed

The acquisition brought 14 XetHub employees into Hugging Face's workforce^[3]. It also fit a period of expansion: Hugging Face had raised a $235 million Series D a year earlier, and in June 2024 it bought the developer tool startup Argilla for $10 million^[3].

For users of the Hub, the significance is mostly infrastructural. The models, datasets and Spaces people upload all sit on top of a storage layer that has to keep working as repositories grow, and moving that layer to a Xet-based backend is a behind-the-scenes change that touches every upload and download on the platform. The team behind it had spent years on machine learning infrastructure, first inside Apple and then at a Seattle startup, and the plan was to apply that work to one of the largest public collections of open models and datasets^[1][3].