ModelRefs / Model Hub — AI Glossary

Model Hub — AI Glossary

A centralized repository for storing, versioning, and sharing trained ML model artifacts.

Overview

Model hubs (HuggingFace Hub, MLflow Model Registry, W&B Artifacts, Vertex AI Model Registry) provide: model versioning, metadata (training config, metrics, data lineage), artifact storage, access control, and deployment integrations. Foundational infrastructure for reproducible ML and organizational model governance.

Reference details

Topic	infrastructure
Last reviewed	2026-06-24

How it works

Model artefacts are stored as versioned, addressable objects, so a specific checkpoint can be fetched by identifier and revision rather than by filename.
Metadata travels with the artefact — training configuration, evaluation results, licence, data lineage — which is what makes a download auditable later.
Clients resolve and cache by revision, so a build pinned to a revision is reproducible and one tracking a branch is not.
Access control gates who may read or publish, which is the difference between a public distribution point and an internal one.
Deployment integrations read straight from the hub, so a serving stack pulls the same artefact the registry recorded.

Example: Why pinning a revision is the whole reproducibility story

A pipeline fetches a model by name with no revision. The upstream author pushes a new checkpoint; the next build silently gets different weights and evaluation moves for reasons nothing in the diff explains. Pinning the revision makes that impossible and makes the upgrade an explicit commit. This is the same discipline as a lockfile for dependencies, and it is skipped about as often — with the difference that a model change alters behaviour rather than breaking the build loudly.

Commonly confused with

The generic category — a repository for sharing and versioning model artifacts — of which the Hugging Face Hub is the largest public instance. It differs from a model registry in audience and purpose: a hub distributes models to other people, a registry governs which internal version is promoted to production. Some products do both, which is why the terms get used interchangeably when they should not be.

When to use it

Reach for it when:

Distributing or consuming model artefacts where versions and metadata must travel together.
Reproducibility requires pinning an exact revision rather than a moving name.
An organisation needs one controlled place models are published and pulled from.

Reach for something else when:

Only calling a hosted API, where there is no artefact to version.
Governing an internal approval lifecycle, which is a registry's job rather than a distribution hub's.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Model Hub — AI Glossary.

Frequently asked questions

What is Model Hub?