Parameters

Parameters are the numbers a model is made of. During training they are adjusted, trillions of times, until the network gets good at next-token prediction; after that they're frozen. When you hear '8B' or '70B', that's the parameter count in billions. Frontier models from Anthropic, OpenAI and Google don't publish theirs, but they're far larger.

Everything the model 'knows' without being told is stored across these numbers, which is what parametric knowledge means. There is no lookup table of facts inside; there is a giant tangle of weights that happens to produce the right continuation most of the time. That's why a model can be fluent about a library and still wrong about its current API: the weights encode a blurred average of what it read up to its knowledge cutoff.

Parameter count matters in two practical ways. It sets how much memory the model needs, which is why a 7B model runs on a laptop and a 400B one doesn't. And it roughly tracks capability, though training quality and effort settings matter just as much. A session never rewrites the parameters; every correction you make lives in context, and the weights stay exactly as shipped.