In my experience, to get the best performance out of different models, they need slightly different prompting.