Neural networks
Parameters are all the numbers a model learns during training, mainly its weights; the parameter count is the usual way to describe how big a model is.
You will hear that a model has '7 billion parameters' or '400 billion parameters.' Each parameter is one learned number. Most are weights on connections; a few are extra adjustments called biases. More parameters means more capacity to store patterns, and usually more memory and computing power needed to run it.
Bigger is not automatically better. A model with more parameters needs more training data to fill that capacity, or it overfits. Research in 2022 showed that many large models were undertrained, and that smaller models fed more data can match them. Design and data quality matter as much as size.
Parameters are set by training and then frozen. They are different from the settings you choose when using a model, like temperature, which are options rather than learned values. As of 2026, the biggest models are believed to have over a trillion parameters.
A phone-sized assistant model might have 3 billion parameters and run offline, while the flagship model behind a cloud chatbot has hundreds of billions.