We study Gibbs partition models, also known as composition schemes. Our main results comprehensively describe their phase diagram, including a phase transition from the convergent case described in Stufler (2018, Random Structures \& Algorithms) to a new dense regime characterized by a linear number of components with fluctuations of smaller order quantified by an $α$-stable law for $1< α\le 2$. We prove a functional scaling limit for a process whose jumps correspond to the component sizes and discuss applications to extremal component sizes. At the transition we observe a mixture of the two asymptotic shapes. We also treat extended composition schemes and prove a local limit theorem in a dilute regime with the limiting law being related to an $α$-stable law for $0< α< 1$. We describe the asymptotic size of the largest components via a point process limit.