A model is a catalog resource whose artifacts live in the git repository of your code, under a directory of their own, pinned to a commit by its training. A transform binds one with Model(...) to load it, conventionally as model=, and the lineage follows: datasets -> train -> model -> transform -> dataset.

The model object

A Model(...) binding, and the Output of a training, arrive as a TransformModel:

Scoring on a schedule

A scoring transform is a transform like any other: build it from the Atlas interface, chain it after the datasets it reads through the lineage, or put it on a schedule. A new version of the model, committed by a training, is picked up by the next build.
The /models endpoints of the catalog are not exposed through the public gateway yet. model.info() and model.update() work from inside a build, but not from a local script. model.path never needs an API call.