Active automata learning gains increasing interest since it gives an insight into the behavior of a black-box system. A crucial drawback of the frequently used learning algorithms based on Angluin’s $$L^*$$L∗ is that they become impractical if systems with a large input/output alphabet are learned. Previous work suggested to circumvent this problem by abstracting the input alphabet and the observed outputs. However, abstraction could introduce non-deterministic behavior. Already existing active automata learning algorithms for observable non-deterministic systems learn larger models if outputs are only observable after certain input/output sequences. In this paper, we introduce an abstraction scheme that merges akin states. Hence, we learn a more generic behavioral model of a black-box system. Furthermore, we evaluate our algorithm in a practical case study. In this case study, we learn the behavior of five different Message Queuing Telemetry Transport (mqtt) brokers interacting with multiple clients.