ModelDelta — Git-style Checkpoint Storage for Fine-tunes

Manage fine-tuned model versions like Git, storing only weight deltas instead of full copies

Problem

When fine-tuning, multiple checkpoints from the same base model quickly fill up disk. Each version is a full model copy, making storage expensive, pulls slow, and sharing painful. Developers need a Git-like version management approach that stores only the weight difference from the base model.

Solution

A local tool plus remote repository that stores fine-tune checkpoints as deltas from the same base model, producing tiny .wdelta files. It supports version browsing, rollback, export to standard formats, and push to remote repositories for sharing, cutting storage and transfer costs.

Users

Individual researchers and small AI teams doing model fine-tuning, especially those managing multiple checkpoints

MVP

Support Safetensors/PyTorch; compute and compress weight deltas into .wdelta from the same base model; CLI for save, load, rollback; local version index; export to standard format; optional remote storage

Revenue

Open-source CLI; remote storage billed by volume (e.g., $0.10/GB/month); team subscription per seat

Why now

Fine-tuned models are getting smaller and more common, but checkpoint management is still primitive; indie developers can start with a lightweight CLI and local delta engine, then add remote collaboration