Xet-Backed Repository Storage

Repository uploads and downloads now run over Xet, which deduplicates at the chunk level rather than the file level.

For the way model repositories actually evolve — a checkpoint reuploaded with a handful of tensors changed, a dataset shard corrected in place — this is the difference between transferring the whole file again and transferring the part that moved. Large repositories see the biggest gain, because the unchanged bulk is never sent twice.

Nothing changes in how you push or pull. git push, huggingface_hub, and the MEGA CLI use Xet transparently where the repository is backed by it, and fall back to LFS where it is not.

Transfers are recorded per attempt so a stalled or retried download is visible in your repository's traffic view rather than disappearing into an aggregate.