Three kinds of different age rank milk powder are identified by using the near infrared spectroscopy with support vector machines (SVM). First, the Kennard-Stone method is used to select 120 training sets from 150 samples and other 30 samples are used as the prediction sets. In the experiment, the radial basis function is selected as the kernel function and the two-step grid searching and five-fold cross validation are used to optimize two model parameters: kernel function γ and penalty factor C. The optimal γ and C are 0.03125 and 2048 respectively. The correction model established with the optimal parameters has an identification rate of 100% for both training sets and prediction sets. By comparison with the principal component analysis (PCA), the SVM exhibits its higher identification accuracy. This shows that the near infrared spectroscopy can identify the varieties of different age rank milk powder quickly and accurately.