parameters in an LLM

Parameters are the model's learned numbers — the dials that training turns. A modern LLM has billions of them, and together they encode everything it knows: grammar, facts, writing styles, coding patterns. When people say a model is 70 billion parameters, they are describing how many of these adjustable numbers it carries.

Each parameter is just a weight that scales a signal as text flows through the network. Training nudges every one of them, over and over, so that the model's next-token predictions match real text better. Knowledge is not stored in a lookup table you can read; it is smeared across all the parameters at once. More parameters generally mean more capacity to learn, at the cost of more memory and compute.