Rainfall-runoff models are useful tools for hydrological research, water engineering and environmental applications. Given the large number of available rainfall-runoff models, many comparative studies have been done to compare models performances, to specify their domain of application and provide guidance to end-users. Although most existing comparative studies tested models only on a few basins, we believe that effective model evaluation requires large samples of test catchments. Large test samples however raise the issue of appropriate criteria to quantify model performances. This paper shows that the widely used Nash and Sutcliffe criterion may be difficult to apply for large test samples and that a bounded version of this criterion (called C2M) is better suited for extensive model assessment.