Abstract. Targeting a long-term effort towards a global weather and climate model with a local refinement function, this study systematically configures and evaluates the performance of an unstructured model based on the variable-resolution (VR) approach. Aided by the idealized dry- and moist-atmosphere tests, the model performance is examined in an intermediate degree of complexity. The dry baroclinic wave simulations suggest that the 3D VR-model can reproduce comparable solutions in the refined regions as a fine-resolution quasi-uniform (QU) mesh model, although the global errors increase. The variation of the mesh resolution in the transition zone does not adversely affect the wave pattern. In the coarse-resolution area, the VR model simulates a similar wave distribution to the low-resolution QU model. Two multi-region refinement approaches, including the hierarchical and polycentric refinement modes, further testify the model performance under a more challenging environment. The moist idealized tropical cyclone test further enables us to examine the model ability in terms of resolving fine-scale structures. It is found that the VR model can have the tropical cyclone stably pass the transition zone in various configurations. A series of sensitivity tests examines the model performance in a hierarchical refinement mode, and the solutions exhibit consistency even when the VR mesh is slightly perturbed by one of the three parameters that control the density function. Moreover, only the finest resolution has a dominant impact on the fine-scale structures in the refined region. The tropical cyclone, starting from the 2nd-refinement region and passing through the inner transition zone, gets intensified and possesses a smaller area coverage in the refined regions, as compared to the QU-mesh model that has the same number of grid points. Such variations are consistent with the behavior that one may observe when uniformly refining the QU-mesh model. Besides the horizontal resolution, the intensity of the tropical cyclone is also influenced by the Smagorinsky horizontal diffusion coefficient. The VR model exhibits higher sensitivity in this regard, suggesting the importance of parameter tuning and proper model configurations.