From first to mature versions of Model-Driven Development (MDD) tools, there is a gap, as for any other software applications. All functional requirements must be met, including qualities of services, at the risk of seeing MDD tools rejected by users. In this paper, we focus on performance, especially for large-scale developments. After an overview of methodological elements, we give a list of reusable practices on performance. We conclude by a set of observations and stakes in order to understand where efforts must be applied during the development process.