Numerical values inside a model that determine its behaviour, adjusted during training to encode learned patterns.
Also known as: weights, model weights, weights and biases
In plain English
Parameters are like millions of tiny knobs inside an AI that change how it responds. During training the learning process turns those knobs so the model gives useful answers. Once trained, the knobs stay fixed unless you retrain or fine-tune the model.
In practice
When selecting or deploying models, parameter count often correlates with capability but also with cost: more parameters usually require more memory, compute and higher inference latency. Teams weigh parameter size when choosing hosted large models versus smaller on‑device models, or when using fine‑tuning and adapters to adapt behaviour without full retraining.
Under the hood
Parameters are the model's learned weights and biases—stored as tensors (commonly float32/16 or quantised formats)—updated by optimisation algorithms (e.g. SGD or Adam) to minimise a loss over training data. Parameter count determines model capacity and memory footprint; during inference parameters are fixed, though techniques like fine‑tuning, low‑rank adapters or quantisation modify or compress them.
Example
We should choose a model with fewer parameters to meet the GPU memory budget and lower inference cost.
Related terms
Neural Network · AI Training · Fine-tuning · Inference · Foundation Model