Hexz brings versioned, deduplicated storage for ML checkpoints to your terminal. Open source and self-hosted.
pip install hexzInitialize repos, commit checkpoints, diff snapshots, and mount weights for serving — all from one binary.
Create a weight repository backed by your object store.
$ hexz init s3://ml-checkpoints/llm-7b
Initialized empty weight repository
remote: s3://ml-checkpoints/llm-7b
config: .hexz/config.tomlWeights are chunked and addressed by hash. Identical data is stored once regardless of how many snapshots reference it.
Each commit records references to chunks, not copies. New checkpoints only write the bytes that actually changed.
Read individual tensors or slices without downloading full checkpoint files. Useful for serving and evaluation.
Backends plug into S3, GCS, or compatible APIs. No proprietary service or mandatory control plane.
Tag, diff, and inspect any snapshot. Know exactly which weights backed a deployment or evaluation run.
No runtime dependencies, no daemon, no cluster. Download the binary and start versioning weights.
Early access is limited. Leave your email and we'll share progress, invite a small group of teams to shape the roadmap, and let you know when the CLI is ready.