A trained machine learning system that maps inputs to outputs using learned parameters.
What It Is
In this reference, a model is the end product you run at inference time. It includes learned parameters and the forward pass that applies them. Checkpoints, Hugging Face repos, and API endpoints all refer to concrete model instances, not abstract design patterns.
Why It Matters
Separating model from architecture, module, and component helps you compare releases, trace bugs, and search the atlas without conflating a family name with a specific trained checkpoint.
Simple Example
GPT-2 124M is a model: a particular set of weights trained with a transformer architecture. The architecture page explains how layers connect; the model page names the artifact you load or call.
Common Confusions
A model is not the same as an architecture: many checkpoints can share one architecture. A model is also not a single layer or attention block—those are modules and components inside the architecture.