What are parameters in an AI model?

Parameters are the adjustable internal numbers inside a model that get set during training, often numbering in the billions. Each one is small and meaningless alone, a weight nudging one calculation slightly up or down, but together they encode everything the model appears to “know”: grammar, facts absorbed from training text, and the statistical patterns that let it continue a sentence plausibly. When people describe a model as large or small, they are usually describing how many of these numbers it has.

In practice

Training is the process of adjusting parameters so the model gets better at predicting text: given a sequence of words, it learns to assign a higher likelihood to the word that actually comes next. This happens gradually, over huge amounts of text, using an error signal that tells the process which numbers to shift and by how much. Researchers at institutions such as Stanford HAI note that scaling up parameter counts has generally tracked with broader capability, but the relationship is not simple or guaranteed. More parameters raise cost and can produce uneven returns depending on training data quality and architecture, and a model with fewer parameters can outperform a larger one if it was trained more effectively. Whether the resulting capability amounts to genuine “understanding” is itself contested among researchers, and framing bodies like NIST tend to describe model behaviour in terms of measurable performance rather than claims about comprehension.

Not to be confused with

Parameters are learned once, during training, and then fixed until the model is retrained. This is different from the settings a user or developer adjusts at the moment of use, such as temperature, which controls how predictable or varied a response is. Products like ChatGPT, Claude, Gemini and Copilot all expose some of these use-time settings, but none of them let a user touch the underlying parameters directly.