Model Registry

A model registry is the central versioned store that holds every trained model an organisation intends to run, together with the metadata, lineage, and lifecycle state needed to operate it safely. A registered model has a version (v1.0, v1.1, v2.0), a stage (Staging / Production / Archived), a model card, links to the training run that produced it, evaluation metrics from that run, the code and data versions used, and the deployment targets that currently serve it. The registry is the source of truth that ties training-time artifacts to runtime behaviour — without it, teams lose track of which model is currently in production, why it was promoted, and what changed between versions.

The 2026 landscape is dominated by a handful of registry products. **MLflow Model Registry** is the open-source default, integrated with MLflow Tracking, sufficient for most APAC mid-market teams. **Vertex AI Model Registry** and **SageMaker Model Registry** are the Google and AWS managed options, best when the team already lives in that cloud's ML stack. **Weights & Biases Artifacts** and **Neptune** provide registry features as part of their broader ML platform. **Databricks Unity Catalog** has absorbed model-registry features for Databricks-native teams. **Hugging Face Hub** serves as the public registry for open-weight models and increasingly as a private registry for internal models.

For APAC mid-market teams, the practical discipline is **one registry per company, not per team**. A model registered in team A must be discoverable by team B; shared standards for metadata (owner, purpose, license, training-data classification, evaluation coverage) let governance and audit functions operate across the organisation. Integrate the registry with deployment so promotion to Production stage triggers the actual deployment (MLflow + Argo CD, SageMaker Registry + Pipelines, Vertex AI Registry + Endpoints), rather than leaving the registry as a parallel-documentation artifact.

The non-obvious failure mode is **a registry with no gates**. Anyone can promote any model version to Production stage, no review is required, and the lifecycle-stage field becomes decorative rather than authoritative. Configure the registry to require approval (with named approvers), link promotion to model-card completeness and evaluation-threshold checks, and log every stage transition with reason text. A registry where 'Production' means 'someone clicked a button' has all the overhead of governance with none of the benefit.