Model transformation is the key to model-based software engineering. When the model transformation is applied to industrial developments, its scalability becomes an important issue. To test the performance of model transformations, developers must be able to generate a set of models, i.e. the test inputs, efficiently. This paper proposes a randomized approach to generating large models. This approach can produce a model randomly and correctly based on the definition of metamodel and user-defined constraints. And the evaluation result also shows that the proposed approach is more efficient than other approaches, and therefore is more suitable for supporting performance testing of transformations.