Model Registry

A model registry is the centralized catalog of trained machine-learning models — including LLMs, LoRA adapters, embedding models, classifiers, and rerankers — that tracks versions, lineage, metrics, approval status, and deployment state across an organization's AI estate. The registry is the system of record answering "which model is in production for which workload, who approved it, what data trained it, what evaluations passed, and what does the audit need to know about it." Leading commercial and open-source registries include MLflow (the dominant open-source choice, originally from Databricks), Weights and Biases Model Registry, Hugging Face Hub (the de facto registry for open-weight models), AWS SageMaker Model Registry, Azure ML Model Registry, Google Vertex AI Model Registry, and Comet ML. Modern registries support model cards, evaluation results, lineage to training data and hyperparameters, A/B test results, deployment endpoints, and approval workflows. A practical recipe: every model — including fine-tunes, adapters, and provider-hosted models you proxy to — gets a registered version with its hash, training config, eval suite results, owner, approval status, and deployment targets; promotion from "staging" to "production" requires a documented approver and an updated model card; rollback to a prior version is a single registry operation. For LLM-specific registries, version everything: base model + version, fine-tune commit, prompt templates, embedding model, retrieval index version, evaluation suite. The "model" in production is really a stack, and the registry must capture all of it. AI governance teams treat the registry as their single source of truth for AI inventory — the EU AI Act's high-risk system registry, ISO 42001 documentation, and NIST AI RMF Map function all feed from registry exports.

Registry discipline from 25 years of content versioning: Centralpoint's content registry — versioning, approval workflows, audit trails, deployment targets — has governed enterprise content for 25 years and now extends naturally to AI models, prompts, skills, and adapters. The registry stays on-premise, tokens meter per skill, and registry-governed chatbots deploy through one line of JavaScript.


Related Keywords:
Model Registry,Model Registry,Oxcyon, AI, AI Governance, Generative AI, Inference, Inference, Inferencing, RAG, Prompts, Skills Manager,