A model registry is a centralized system for managing trained machine-learning models across their lifecycle, in the same spirit that a code repository manages source code or an artifact repository manages software builds. Instead of trained models living as loose files on someone’s laptop, the registry gives each model a name, tracks every version, and records how each version was produced.
The MLflow Model Registry, a widely used open-source example, describes itself as a centralized model store with APIs and a UI to collaboratively manage a model’s full lifecycle. Typical capabilities include automatic version tracking, lineage that links each model version back to the experiment run, data, and parameters that created it, mutable aliases or stages (such as marking one version as the champion or promoting it to production), and tags and access controls for governance and auditability. These features let teams compare versions, roll back to an earlier model, and know exactly which model is serving traffic.
The model registry is a core piece of MLOps infrastructure, sitting between experimentation and deployment and underpinning reproducibility and governance requirements.
For a business reader, a model registry answers basic but critical questions that grow urgent at scale: which model is in production, how was it built, and can we safely roll it back.