Parameter

Parameters are the dials a model learns. Training is the process of nudging billions of them until the model’s output matches the patterns in its training data. When you read that a model has 276 billion parameters, that is the count of those dials.

More parameters generally means more capacity and a more expensive model to run, which is why the number is quoted so often. It is a poor standalone measure of quality, though. Architecture, training data and how many are actually used per request matter at least as much, and a well-built smaller model routinely beats a larger sloppy one.