Model architecture describes how a neural network's internal components are arranged and how they connect to and process information from one another. It's the blueprint that determines what kind of data a model can handle well and how it processes that data internally, rather than the specific values it learns during training.
Different architectures suit different problems. Convolutional networks are built to recognize spatial patterns and dominate image-related tasks; recurrent networks process sequences step by step and were the standard choice for text before largely being displaced; and transformers, built around the attention mechanism, now underlie essentially every major large language model because they process entire sequences in parallel and capture long-range relationships between words far more effectively than their predecessors.
