Molecular simulation is one of the most important ways of studying biomolecules. In the last two decades, by combining the molecular simulations with experiments, a number of key features of structure and dynamics of biomolecules have been reflealed. Traditional molecular simulations often use the all-atom model or some coarse grained models. In practical applications, however, these all-atom models and coarse grained models encounter the bottlenecks in accuracy and efficiency, respectively, which hinder their applications to some extent. In reflent years, the multiscale models have attracted much attention in the field of biomolecule simulations. In the multiscale model, the atomistic models and coarse grained models are combined together based on the principle of statistical physics, and thus the bottlenecks encountered in the traditional models can be overcome. The currently available multiscale models can be classified into four categories according to the coupling ways between the all-atom model and coarse gained model. They are 1) hybrid resolution multiscale model, 2) parallel coupling multiscale model, 3) one-way coupling multiscale model, and 4) self-learning multiscale model. All these multiscale strategies have achieved great success in certain aspects in the field of biomolecule simulations, including protein folding, aggregation, and functional motions of many kinds of protein machineries. In this review, we briefly introduce the above-mentioned four multiscale strategies, and the examples of their applications. We also discuss the limitations and advantages, as well as the application scopes of these multiscale methods. The directions for future work on improving these multiscale models are also suggested. Finally, a summary and some prospects are preflented.