RegCM3 (REGional Climate Model) simulations of precipitation in China in 1991 and 1998 are very sensitive to the cumulus parameterization. Among the four schemes available, none has superior skills over the whole of China, but each captures certain observed signals in distinct regions. The Grell scheme with the Fritsch-Chappell closure produces the smallest biases over the North; the Grell scheme with the Arakawa-Schubert closure performs the best over the southeast of 100°E; the Anthes-Kuo scheme is superior over the northeast; and the Emanuel scheme is more realistic over the southwest of 100°E and along the Yangtze River Basin. These differences indicate a strong degree of independence and complementarity between the parameterizations. As such, an ensemble is developed from the four schemes, whose relative contributions or weights are optimized locally to yield overall minimum root-mean-square errors from observed daily precipitation. The skill gain is evaluated by applying the identical distribution of the weights in a different period. It is shown that the ensemble always produces gross biases that are smaller than the individual schemes in both 1991 and 1998. The ensemble, however, cannot eliminate the large rainfall deficits over the southwest of 100°E and along the Yangtze River Basin that are systematic across all schemes. Further improvements can be made by a super-ensemble based on more cumulus schemes and/or multiple models.