I mean you can continue to evolve the model weights but the performance would suck so we don't do it. Models are built to an optimal state for a general set of benchmarks, and weights are frozen in that state.