They’re the output of all that data centre processing. The model is fixed at that point and can’t “learn” anything more, so the pressure is always on to release a newer model that resulted from more data with better algorithms, but still small enough t9 be able to run on a consumer grade GPU