Parameters
The millions of internal “dials” where everything learned lives.
The analogy
Picture a mixing desk with billions of tiny dials. During training, each example nudges some dials until the “song” sounds right. What the model learned isn't stored as sentences: it's stored as the exact position of all those dials.
In detail
Parameters (or weights) are the neural network's adjustable numbers; they're set during training and determine how each input is transformed. “7B” or “70B” mean billions of parameters. More parameters usually mean more capability, but also higher training and inference costs; data quality matters as much as size.
An example
A “7B” model fits on a powerful laptop and answers fast; one with hundreds of billions needs specialized servers. That's why model families come in different sizes for different jobs.