Model registry
Weights are artefacts. Treat them like it.
Content-addressed storage, resumable multi-hundred-gigabyte transfers, deduplication across fine-tunes, and a signed provenance record that travels with the file.
Built for files that do not fit in a git repo
Chunk-level dedup
A LoRA on top of a 70B base stores the delta, not the base. Across our registry that is a 2.8× saving and a much faster pull.
Transfers that survive
Resumable, parallel, verified per chunk. A 400 GB pull that dies at 92% resumes at 92%.
Provenance that verifies
Training data manifest, base model, hyperparameters and evaluation results, signed by the publisher's key.
Licences that travel
SPDX identifiers and use restrictions in the manifest, surfaced by the CLI before a pull, not buried in a card.
Safe formats only
safetensors and GGUF. Pickle-based checkpoints are rejected at push because they execute code on load.
Mirrored
Four regions plus a torrent swarm for the largest public releases. Pulls pick the fastest source.
Deduplication
Why the registry is smaller than its contents
Most published models are derivatives. Chunking at a content boundary rather than a file boundary means shared tensors are stored once, regardless of how the shards were split.
- 11.8 PB logical, 4.2 PB physical
- A typical fine-tune adds 0.4–3% new bytes
- Pulls reuse chunks already on your disk from other models
- Verification is per chunk, so corruption is local and cheap to fix