Hugging Face — AI Glossary

Reference details

Related terms

• Model Hub
• Model Registry
• Open Weights

How it works

1. Models, datasets and demos are stored as git repositories with large-file support, so an artefact has a history and a resolvable revision.
2. Client libraries resolve a repository id and revision, then cache locally — which is why a pinned revision is reproducible and an unpinned one is not.
3. Standard configuration files let a model be loaded by a common interface regardless of who published it, which is what made the ecosystem interoperable.
4. Hosted inference and spaces run models on the platform, so a checkpoint can be tried without local hardware.
5. Anyone can publish, so provenance, licence and evaluation claims are the publisher's and need checking rather than assuming.

Example: Why the revision matters more than the name

A pipeline loads a model by repository name with no revision. The publisher pushes a new checkpoint; the next environment build downloads different weights, and evaluation moves with nothing in the diff to explain it. Pinning the revision makes that upgrade an explicit commit instead of a surprise. The discipline is identical to a dependency lockfile, with the difference that the failure is a quiet behaviour change rather than a loud build error.

Commonly confused with

The name covers three things that are worth separating: the company and its ecosystem, the Hub where models and datasets are hosted, and the Transformers library that most open-weight code imports. A model hub is the generic category the Hub belongs to — MLflow's registry and cloud registries are others. Saying "we use Hugging Face" usually means the library, the Hub or the Inference Endpoints product, and those are different dependencies.

When to use it

Reach for it when:

• Obtaining or publishing open model weights and datasets with versioning.
• Loading models through a common interface rather than per-publisher code.
• Reproducibility requires pinning an exact revision.

Reach for something else when:

• You only call a hosted proprietary API, where there is no artefact to fetch.
• An internal approval lifecycle is needed, which is a registry's job rather than a distribution hub's.
• Trusting publisher claims — licence and provenance need verifying independently.

Continue your research

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

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path