When the new pre-trained parameters come out in a new model generation, your old fine tuning doesn't apply to them.