Durum wheat is a crop of great economic relevance for Mediterranean regions, especially in developing countries. A decreasing trend in durum production is expected in the near future because of several factors, in particular environmental constraints due to climate change and variability. The aim of this work was to test the predictive performance of CERES-Wheat model, implemented in DSSAT software systems, under Mediterranean climate condition and soil types of Southern Sardinia, Italy. CERES-Wheat model was calibrated for three durum wheat Italian varieties (Creso, Duilio, and Simeto) using a 30-year data set (1974–2004) and a trial and error iterative procedure. Then, the model was validated and evaluated using several statistics. The model showed a quite good performance in predicting grain yield and anthesis date, with errors comparable with those reported by other studies conducted on bread and durum wheat. Predictions of grain weight and grain number did not match very well observations, confirming the difficulties of CERES-Wheat in estimating grain yield components. The results of this study suggest the need of specific field experiments and further model evaluations and improvements to better understand model simulation results of grain yield components of durum wheat.