Abstract Regional weather hazards need local detail but remain constrained by the surrounding global circulation. Current AI weather forecasting mainly relies on globally uniform latitude‐longitude tensors, spherical graphs, or limited‐area models with external boundary forcing. Numerical weather prediction offers another route through a stretched cubed sphere (SCS) mesh, which keeps the globe closed while refining a target region. Motivated by this, we develop one‐step and multistep StretchCast models on the SCS mesh, with 7,776 global cells and about 0.875° ${}^{circ}$ mean resolution over eastern China. The one‐step model establishes stable multivariate prediction and shows that local facewise mixing and global coupling improve forecast quality. The multistep model improves medium‐range forecasting performance, preserves continuity for Typhoon Muifa across a face boundary, and retains broadly consistent spectral structure. These results identify the SCS mesh as a promising route to global‐regional AI weather forecasting.