A model registry typically stores: the model artifact (weights, configuration), training metadata (hyperparameters, dataset version, training duration), evaluation metrics (accuracy, latency, fairness metrics across demographics), deployment status (which version is serving in production), and lineage (which experiment, code commit, and data pipeline produced this model). MLflow Model Registry, Weights & Biases, and SageMaker Model Registry are popular implementations.
In production workflows, the model registry is the handoff point between training and serving: a data scientist trains and evaluates models, registers the best one, a reviewer approves it, and the deployment system pulls the approved model and serves it. This separation of concerns — training doesn't directly touch production, deployment only uses registry-approved models — reduces the risk of deploying broken models.
LLM registries have specific needs: models are very large (tens to hundreds of GB), fine-tuned variants share a common base model (store adapters separately), and evaluation is more complex (automated benchmarks + human evaluation + safety checks). Hugging Face Hub serves as a de facto model registry for the open-source community, with model cards, versioning, and evaluation results. Enterprise teams often use private registries for proprietary models.