Model Architecture
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
The structural design of a machine learning model: its layers, connections, and operations. The blueprint that training fills in with learned weights.
What is Model Architecture?
Model architecture is the structural design of a machine learning model: which layers it has, how they connect, and what operations transform input into output.
The architecture is a blueprint, not the finished thing. It fixes what is learnable; training then fills in the weights; the result is a model. Two teams using the same architecture on different data end up with different models, the way two houses built from one blueprint hold different furniture.
Architecture, Model, and Weights
The three terms get conflated, and the distinction matters when reading a system card or an audit. The architecture is the design: layer types, sizes, and connections, chosen by engineers before training. The weights are the millions or billions of numbers that training assigns to that design.
The model is the trained artifact: architecture plus weights, the file you actually deploy. “Llama-style architecture” describes a design anyone could train; a specific released model is one set of weights in that design.
Types of Model Architecture
- Feedforward network: Data flows one way from input to output, the baseline design for tabular prediction tasks.
- Convolutional Neural Network (CNN): Sliding filters extract spatial features from grid data such as images.
- Recurrent Neural Network (RNN): Loops carry a hidden state across time steps for sequential data.
- Transformer: Self-attention processes whole sequences in parallel; the design behind modern language models.
- Graph Neural Network (GNN): Message passing between connected nodes handles graph-structured data.
The pattern across all five: architecture choice follows data shape. Grids suggest CNNs, sequences suggest RNNs or transformers, graphs suggest GNNs.
Example of Model Architecture
A team building a skin-lesion classifier writes down an architecture before any training: an input layer for 224x224-pixel images, five convolutional layers to extract visual features, pooling layers between them, and two fully connected layers ending in a benign/malignant output.
At this point the design predicts nothing; its weights are random. After training on labeled dermatology images, the same architecture carries learned weights and becomes a usable model. When the hospital later retrains on newer images, the architecture stays identical – only the weights, and therefore the model, change.
Related AI terms: Model · Weights · Convolutional Neural Network · NLP Transformer · Wide & Deep
Did you like the Model Architecture gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About