Hugging Face buys XetHub to kill Git LFS on its 12PB model hub

Hugging Face is acquiring XetHub, a Seattle startup founded by three ex-Apple engineers who built Apple’s internal ML infrastructure. The deal isn’t about folding in another AI tool — it’s about replacing the storage backbone of the entire Hugging Face Hub, which currently hosts 1.3 million models, 450,000 datasets, and 680,000 Spaces. The company’s CTO, Julien Chaumond, said the XetHub team will help unlock the next five years of growth by switching the Hub to a better version of Git LFS. That’s a candid admission that Git LFS was always a stopgap: fine for bootstrapping in 2020, but never designed for the multi-gigabyte model files and Parquet datasets that now define AI development.

XetHub’s core technology is chunked file storage with deduplication. The practical difference is stark. Today, adding a single row to a 10GB Parquet file means re-uploading the whole thing. Under the XetHub approach, only the few chunks containing that new row need to move. The same logic applies to GGUF model files — updating one metadata value in a Llama 3.1 405B repo could mean re-uploading a few kilobytes instead of the entire model. For a platform pushing 6PB of daily Cloudfront bandwidth and handling a billion requests per day, that efficiency translates into real infrastructure savings and faster iteration for the community.

The founders bring serious pedigree. Yucheng Low, Ajit Banerjee, and Rajat Arya all worked on Apple’s internal ML stack, and Low previously founded GraphLab (later Dato and Turi), which Apple acquired. In his announcement, Low noted he has worked in the data domain since his PhD and scaled AI data management past 100PB at Apple, supporting dozens of teams shipping hundreds of features annually. His framing of the acquisition centers on a lesson he says he underestimated: the power of data. What seemed impossible a few years ago became solvable with orders of magnitude more data and models large enough to absorb it.

The integration will also tackle collaboration problems that enterprise teams face when working with massive datasets and models. XetHub was founded in 2021 with Madrona backing specifically to make ML teams operate like software teams — a goal that aligns neatly with Hugging Face’s push into enterprise. The announcement hints at features for understanding how datasets evolve over time, though details remain thin. With trillion-parameter models already appearing in the community, the timing makes sense. Git LFS was never built for this scale, and Hugging Face is betting that a custom storage backend is the only way to keep growing without collapsing under its own bandwidth bills.