Large language models can have billions, or even trillions, of parameters. But how big do they need to be to achieve acceptable performance?