Species distribution models (SDMs) have been increasingly used to predict the geographic distribution of a wide range of organisms; however, relatively fewer research efforts have concentrated on rare species despite their critical roles in biological conservation. The present study tested whether community data may improve modelling rare species by sharing information among common and rare ones. We chose six SDMs that treat community data in different ways, including two traditional single-species models (random forest and artificial neural network) and four joint species distribution models that incorporate species associations implicitly (multivariate random forest and multi-response artificial neural network) or explicitly (hierarchical modelling of species communities and generalized joint attribute model). In addition, we evaluated two approaches of data arrangement, species filtering and conditional prediction, to enhance the selected models. The model predictions were tested using cross validation based on empirical data collected from marine fisheries surveys, and the effects of community data were evaluated by comparing models for six selected rare species. The results demonstrated that the community data improved the predictions of rare species' distributions to certain extent but might also be unhelpful in some cases. The rare species could be appropriately predicted in terms of occurrence, whereas their abundance tended to be underestimated by most models. Species filtering and conditional predictions substantially benefited the predictive performances of multiple- and single-species models, respectively. We conclude that both the modelling algorithms and community data need to be carefully selected in order to deliver improvement in modelling rare species. The study highlights the opportunity and challenges to improve prediction of rare species' distribution by making the most of community data.