Abstract: In this experiment, spectrum of different parities milk samples were scanned using near-infrared spectrometer (NIRS) , and the protein content in milk samples was measured using the multifunctional analyzer of dairy. The models were established through orthogonal design and validated with principle component regression (PCR) , partial least square (PLS) , modified PLS and a variety of pretreatment methods. Then, the models were evaluated by the objective function method. The result showed that the first parity samples share the same optimal model with the second parity samples in the milk protein And their calibration coefficient of correlations (R2) , standard error of calibration (SEC) and standard error of prediction (SEP) were 0.9626, 0.0531, 0.0630 and 0.9377, 0.0810, 0.1100, respectively. The optimal model of the milk protein of the third parity samples and more times samples were established, and the R2, SEC and SEP of the optimal model were 0.9406, 0.0461, 0.0500. Meanwhile, the optimal model of all milk protein samples were also established, and the R2, SEC and SEP of the optimal model were 0.9351, 0.0687, 0.0790. Therefore, it was suitable for using these models to detect simulately the milk protein rapidly, accurately and non-destructively. The analysis method provided the theoretical basis for detecting the milk protein quickly in mixed raw milk.