Snow is a granular material consisting of a network of interconnected ice grains, which implies that its mechanical response depends on both the microstructure and the material properties of ice. Multiple numerical approaches can be used to model snow-like granular materials. In the Discrete Element Method, the explicit modeling of each ice grain makes it possible to capture the influence of the microstructure, but computation time can become quite important in large systems. The continuum approaches, such as the Finite Element Method, can reduce the computational cost, but the existing constitutive models do have some limitations in accounting for the microstructure. The current work proposes a multiscale constitutive modelling approach, based on the 3D H-model initially developed for geomaterials [1]. The 3D H-model is a multiscale constitutive material model that accounts for the microstructure of a granular material. The basic principle lies on the description of a Representative Elementary Volume (REV) as a statistical distribution of 3D grain cells with different orientations. The use of a bi-hexagonal structure to describe the grain cells allows for deriving an analytical relationship between the strains and the stresses acting within each cell. The response also includes the complex contact interaction between the grains. To account for the particularities of snow, the contacting grains are initially linked by ice bonds, described as constant volume elasto-viscoplastic beams [2] [3]. When a bond between two grains fails, a residual elasto-frictional contact between those grains applies. The upscaling process allows modeling the snow behavior at the specimen scale (REV). Confined compression tests at the REV scale were numerically performed. A parametric study and comparisons with experimental results from the literature [4] demonstrate the capability of the H-model to reproduce the complex snow behavior in a specific density range.