Model versioning is the practice of assigning distinct identifiers to iterations of a machine learning model so teams can track, compare, and retrieve specific model artifacts. A version may be created when a model is retrained, its architecture changes, or a different set of parameters is registered. Model registries commonly associate each version with metadata such as its source run, creation time, description, tags, and aliases.
For example, a team can keep one registered model name while recording versions for successive training runs. It can then compare evaluation results or identify which artifact was deployed in an environment. Model versioning differs from data versioning, which tracks changes to datasets, and from source-code version control. It also does not prove that a model is approved or safe to release. Those decisions require separate evaluation, qualification, and release evidence. Versioning provides a stable reference for a model artifact, but the surrounding evidence determines whether that version is suitable for its intended use.