Model-based testing (MBT) suffers from two main problems which in many real world systems make MBT impractical: scalability and automatic oracle generation. When no automated oracle is available, or when testing must be performed on actual hardware or a restricted-access network, for example, only a small set of test cases can be executed and evaluated. However, MBT techniques usually generate large sets of test cases when applied to real systems, regardless of the coverage criteria. Therefore, one needs to select a small enough subset of these test cases that have the highest possible fault revealing power. In this paper, we investigate and compare various techniques for rewarding diversity in the selected test cases as a way to increase the likelihood of fault detection. We use a similarity measure defined on the representation of the test cases and use it in several algorithms that aim at maximizing the diversity of test cases. Using an industrial system with actual faults, we found that rewarding diversity leads to higher fault detection compared to the techniques commonly reported in the literature: coverage-based and random selection. Among the investigated algorithms, diversification using Genetic Algorithms is the most cost-effective technique.