We consider learning to play multiplayer imperfect-information games with\nsimultaneous moves and large state-action spaces. Previous attempts to tackle\nsuch challenging games have largely focused on model-free learning methods,\noften requiring hundreds of years of experience to produce competitive agents.\nOur approach is based on model-based planning. We tackle the problem of partial\nobservability by first building an (oracle) planner that has access to the full\nstate of the environment and then distilling the knowledge of the oracle to a\n(follower) agent which is trained to play the imperfect-information game by\nimitating the oracle's choices. We experimentally show that planning with naive\nMonte Carlo tree search does not perform very well in large combinatorial\naction spaces. We therefore propose planning with a fixed-depth tree search and\ndecoupled Thompson sampling for action selection. We show that the planner is\nable to discover efficient playing strategies in the games of Clash Royale and\nPommerman and the follower policy successfully learns to implement them by\ntraining on a few hundred battles.\n