A simplified representation of something, used to make predictions about it. In machine learning it means the trained artefact: the structure plus all the numbers learned from data. When people say ‘the model got it wrong’ they mean this thing, a file of numbers plus the code that runs it, not a program somebody wrote.
Technical view
Comprises an architecture, which defines the computation, and parameters, which are learned. Serialised as weight files in formats such as safetensors, often accompanied by configuration, tokeniser and chat template. The word is heavily overloaded: a statistical model, a mental model and a deployed system all get called one, and the ambiguity causes real confusion in governance discussions where ‘the model’ may mean the weights, the served endpoint, or the whole product.
Related
Part of AI Terminology Index, section: The Foundations.