The ability to model granular systems at the level of individual particles has largely conduced to the success of the discrete element method (DEM). At the same time, this fundamental concept hinders the use of the DEM for industrial-scale simulations as the computational cost of the method increases with the size of the system. The DEM coarse-grain (CG) model provides one means of counteracting this effect by replacing a group of original particles by a larger (pseudo) particle. The major shortcoming of this approach is that it fails to capture effects that intrinsically depend on particle size. To overcome this deficiency we have devised a novel model to efficiently combine multiple levels of coarse-graining in a single DEM simulation. While a coarse realization is used where it sufficiently represents the granular flow, the level of resolution may be increased recursively in spatially confined regions of interest. Thus, the method is able to benefit from the speedup of the coarse-grain approach and retain the details of the granular system in crucial regions. Two-way coupling between different levels of resolution is established by passing volume-averaged flow properties. We present validation data based on the comparison between the computed statistical properties of our multi-level coarse-grain (MLCG) model and the corresponding properties of the fully resolved reference system.