Ensembles of weakly fitted randomized models have been studied intensively and used successfully in the supervised learning literature during the last two decades. Among the advantages of these methods, let us quote the improved scalability of the learning algorithm thanks to its randomization and the improved predictive accuracy the induced models thanks to the higher flexibility in terms of bias/variance trade-off. In the present work we propose to explore this idea in the context of density estimation. We propose a new family of unsupervised learning methods of mixtures of large ensembles of randomly generated poly-trees. The specific feature of these methods is their scalability to very large numbers of variables and training instances. We explore these methods empirically on a set of discrete test problems of growing size. We finally discuss possible extensions which we plan to study.