The increasing availability of distributed rainfall data and computational resources is providing the opportunity the use distributed models for rainfall-runoff forecasting or other applications. This paper compares the accuracy of simulations from a complex distributed model (KINEROS), a simple distributed model (based on the Soil Conservation Service (SCS) method), and a simple lumped model (SCS method). The 150 km2, semiarid Walnut Gulch experimental watershed was the test site; models were validated using 24 severe thunderstorms and rain gauge densities similar to those found at flash flood warning sites (one gauge per 20 km2). Under these circumstances, none of the models were able to accurately simulate peak flows of runoff volumes from individual events. Models showed somewhat more skill in predicting time to peak and the ratio of peak flow to volume. When calibration was performed, the accuracy of the complex distributed model was similar to that of the simple distributed model. Without calibration, the complex distributed model was more accurate than the simple distributed model. The spatially lumped model performed very poorly. The complex distributed model was validated under real-time forecasting conditions; forecasts based on observed rainfall had lead times of 30--75 min.