What is model architecture?
Model architecture is the blueprint of a model: which kinds of layers it has, how many, how wide, and how they connect. The architecture is fixed before training and does not change; training fills in the model weights that live inside it. Two models can share an architecture and behave completely differently because they were trained on different data.
A handful of architectures dominate. Convolutional neural networks process images through layers that scan for local patterns. Recurrent networks and their long short-term memory (LSTM) variant process sequences step by step. The transformer, built around the attention mechanism, now underlies nearly all large language models and many vision and speech models. Outside deep learning, a gradient-boosted tree ensemble or a linear regression is also an architecture in the loose sense of a model’s form.
Why model architecture matters
Architecture determines what a model can take as input, how it scales, and what hardware it needs. A transformer’s cost grows with the length of the input, which is why context window limits exist. A convolutional network runs well on a small edge device in a factory camera; a large transformer does not. When a hospital’s imaging team evaluates a vendor’s chest X-ray model, the architecture listed in the model card tells its engineers whether the model will run on the existing on-site graphics processing unit (GPU) and how it will handle images of unusual size. Most buyers never choose an architecture, but they should confirm the one they are buying fits their hardware and data.