Model-Driven Engineering (MDE) exploits well-defined, tool-supported modelling languages and operations applied to models created using these languages. Model transformation is a critical part of the use of MDE. It has been argued that transformations must be engineered systematically, particularly when the languages to which they are applied are large and complicated – e.g., UML 2.x and profiles such as MARTE – and when the transformation logic itself is complex. We present an approach to designing large model transformations for large languages, based on the principle of separation of concerns. Specifically, we define a notion of localized transformations that are restricted to apply to a subset of a modelling language; a composition of localized transformations is then used to satisfy particular MDE objectives, such as the design of very large transformations. We illustrate the use of localized transformations in a concrete example applied to large transformations for system-on-chip co-design.