What are parameters in an AI model?

In plain English

Parameters are the internal numbers a neural network adjusts while it learns. Each connection between units has a weight — a number saying how much influence one signal has on the next — and training is simply the process of nudging all of those numbers until the model’s outputs improve.

After training, the parameters are the model. Everything it has absorbed about grammar, facts or visual shapes is stored in that giant list of numbers. Modern models have billions of them, which is why their files are so large and why you often see sizes quoted as “7B” or “70B” — meaning billions of parameters.

Why it matters

Parameter count is the industry’s rough shorthand for how big and capable a model is. It also has very practical consequences: more parameters generally means more memory needed to run the model, higher costs per answer and slower responses. That is why smaller models are built to run on laptops and phones while the largest live in data centres. When people discuss “open-weight” models, the weights they mean are exactly these parameters.

An example

Think of a mixing desk with billions of sliders. Training is an automated process that inches each slider up or down until the overall sound matches the target. Once finished, nobody adjusts the sliders by hand — the arrangement is the learned knowledge.

What to watch out for

Bigger is not always better. A smaller model trained on higher-quality data, or tuned carefully for a particular job, often beats a much larger general one — and costs far less to run. Parameter counts are also not comparable across different architectures, and many providers no longer publish them. Finally, individual parameters mean nothing on their own; you cannot look inside and find “the fact about Paris”.