Credit scoring model is the core of credit rating model，and the weight is the key factor which influences the accuracy of credit scoring． A rational weight vector is of crucial importance to credit rating． This paper establishes a nonlinear programming which aims at maximizing a standardized mean difference of the credit scores of two kinds of clients． Then by solving the programming，the optimal weight of credit scoring model is derived． To guarantee the effectiveness，the model is validated by Precision Ｒecall Curve． Next，altogether 1231 enterprises are rated based on the credit score calculated by our model． The innovations and characters in this paper are: firstly，an objective function is established by constructing the function relationship between the weight vector W and the standardized mean difference D of the two group of clients． Secondly，a nonlinear programming model，which maximizes the above mentioned objective function，is established to find the optimal weight vector W． This method can guarantee that the scores based on the weights have a significant default discrimination power and can distinguish the default to the largest possible extent，thus avoiding the drawbacks of recent researches． Secondly，non-financial indexes，accounting for nearly 58%，are more important than financial indexes on small enterprise credit rating． Among the non-financial indexes，per capital disposable income of urban residents is most important，with a weighting of 14．7%． Thirdly，the credit scoring model is validated by PＲ curve method． The research result shows that our credit scoring model has a better discrimination power than classic models in existing researches．