Particle clustering has long been a focus in the study of gas–solid flow. Detailed flow field information below the particle scale is required to understand the mechanism of its formation and the statistical properties of its dynamic behavior, but is not easily obtained in both experiments and numerical simulations. In this article, a meshless method is used to reveal such details in the destabilizing of a suspension with hundreds of particles. During the process, doublets, quadruplet and larger clusters are seen to form and disintegrate dynamically, showing a tendency to minimize local voidages. At the same time, single vertical streams, pairs of parallel streams and many irregular streams appear and disappear between particle clusters alternatively, exhibiting a tendency to suffer lowest resistance. Globally, the spatio-temporal compromise between these two tendencies results in a configuration of large clusters separated by fast flow streams. In the clustering process, the inter-phase slip velocity is seen to increase long after the forces on each phase have stabilized, suggesting that inter-phase friction is not a function of local voidage and Reynolds number only, as commonly considered. The article concludes with prospects on the sub-grid scale models for continuum description of gas–solid flow that can be established upon such simulation results.