This study assesses the performance of a regional WRF configuration used operationally by SIMEPAR (SIWRF) and a set of MPAS experiments at 5-km grid spacing over southern Brazil to predict four recent severe weather events. MPAS is tested using a 200-5 km global variable-resolution (VR) mesh with two physics suites—Mesoscale Reference (MR) and Convection-Permitting (CP)—and two sources of initial conditions — NCEP-GFS and ECMWF-IFS forecasts. Skill is evaluated with respect to 48-h accumulated precipitation (correlation, bias, RMSE), hourly precipitation (fractions skill score; FSS, and QPF measures), radar reflectivity (FSS), near-surface variables (MAE, Taylor diagrams), and vertical soundings (bias, RMSE). Across events, MPAS generally outperforms SIWRF for precipitation, with CP physics suite often reducing positive bias relative to MR, while SIWRF tends to underestimate totals during the heaviest events. Notably, MR-IFS achieves the highest correlation for the event dominated by a long-lived, organized convective system, whereas CP-IFS provides the most balanced metrics in two other cases, including the coastal event, consistent with improved placement and intensity of convection. SIWRF exhibits the lowest skill for radar reflectivity, while MPAS shows event-dependent gains, particularly for MR-IFS and CP-GFS. Analysis of convective parameters reveal that MR configurations produce higher median CAPE and 0–3-km SRH than CP and SIWRF, implying more favorable storm environments. These results highlight event-dependent sensitivity to physics and initial conditions and suggest that tailoring MPAS configuration to event features (e.g., favoring CP physics for coastal/diurnally forced convection and MR physics for strongly forced organized systems) may improve forecasts, but this inference is preliminary and requires broader testing (larger case set, data assimilation, ensemble and operational-runtime evaluation) before operational adoption. • First comparison of MPAS and operational WRF for severe weather in southern Brazil. • MPAS with 200-5 km global VR mesh often outperformed 5-km regional WRF. • MPAS precipitation skill and biases strongly depend on physical options and IC source. • MPAS forecasts improved when initialized with ECMWF-IFS ICs rather than NCEP-GFS.