The situation with models right now is that to eke out that last bit of performance, you have to do some things in ways that are specific to the model in question - wording of prompts, when and where to introduce relevant parts into context etc.