Modeling the dynamics of stock returns is very important for derivatives pricing and risk management. Nonnormality is one of the most important stylized facts of stock returns. This paper applies the nonparametric specification test and compares alternative ways of modeling nonnormality in Chinese stock market, including jump models, the models with GED distribution and the models with t distribution. Empirical results show that the models with t distribution could effectively model the stock return dynamics in China.