What is a model registry?
A model registry is a central catalog that stores trained models along with their versions, metadata, evaluation results, and lifecycle stage. Each registered version points to the training run that produced it, the code commit, the data snapshot, the metrics on holdout data, and the person who approved it. Stages such as staging, production, and archived mark where each version sits, and moving between stages can require sign-off.
Deployment systems pull the approved version from the registry rather than from a shared folder or a data scientist’s laptop.
Why a model registry matters
The registry answers the questions that arrive during an incident or an audit: which model made this decision, what was it trained on, who approved it, and how well did it score at the time. A collections agency that scores accounts for likelihood to pay can, when a complaint arrives, pull the exact version that scored that account on that date and show its evaluation record. Rollback becomes a registry operation instead of a search through email. For companies that keep an AI inventory, the registry is usually the system of record for models.
What to watch for
- Register the feature definitions and preprocessing code with the model, or the version cannot be reproduced.
- Require an approval record for promotion to production, even when the approver is on the same team.
- Include downloaded open-weight models and their versions, since they need the same tracking as models trained in-house.