Despite many data clustering methods are available, most of them uncover compactness or connectivity as the intrinsic structure of unlabeled data. Very few approaches explicitly consider the cluster size distribution, especially over‐dispersed (high variance), which may represent yet another important aspect of structural information of unlabeled data. In this paper, we propose a novel joint mixture model framework to estimate cluster size distribution together with cluster compactness (density). Our framework is sufficiently flexible and general to capture a wide range of cluster size distributions from data with mixed feature types. Experiments on clustering synthetic and real‐world data demonstrate a superior performance of our clustering approach in recovering the hidden structure of unlabeled data.