AI glossary · Advanced · MLOps & LLMOps
Model registry
A model registry is a central repository that stores trained AI models by version and records and manages what data and settings each model was built with and what stage it is currently in (under validation, in production, or archived).
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Data, ops & platforms
- Fields
- MLOps & LLMOps
- Added
- 2026-10-08
How it works
After a model is trained, its output files are "registered" in the registry, where they get a name and a version number. Information such as the training data, code version and evaluation scores (metadata) is stored with them. A registered model usually moves through stages such as development → staging (pre-production validation) → production (live use) → archived, and moving between stages can require an approval step. The model serving system pulls the version marked "production" from the registry and deploys it.
Stages a model version goes through
- 1RegistrationModel files + data, code and score records
- 2DevelopmentVersion number assigned
- 3StagingPre-production validation and approval
- 4ProductionPulled by the serving system
- 5ArchivedKept for rollback
Why it matters
When there is only one model, a single file is enough. But as teams and models multiply, no one may know for sure exactly which model is running in production. The registry gives a single answer to that question. If a new model causes problems, you can quickly roll back to the previous version, and for audits or regulatory reviews you can trace which model was deployed, when, and on what basis. In MLOps, it is the link between training and deployment.
Good to know
A registry is a store and a ledger for model files; it does not guarantee model quality. The team has to decide what criteria a model must pass to move to the next stage. Connecting it to an experiment tracking tool makes it easy to trace which experiment each version came from. Even teams that do not train their own LLMs and only use external APIs often version prompts or fine-tuned adapters (such as LoRA) in a similar way.
Example
MLOps tools such as MLflow and the machine learning platforms of major cloud providers include model registry features. For example, a recommendation system team trains a new model, registers it in the registry, validates it with an A/B test in the staging stage, and then promotes it to production. If metrics later get worse, they roll back by setting the previous version in the registry as production again.