As an indispensable drink of people’s daily life, milk’s quality has been also increasingly concerned by consumers. Rapid and accurate detection of milk and its products is the indispensable step for improving the quality of milk and daily products in production. However, traditional methods cannot meet the need. In this paper, rapid quantitative detection of true protein in pure milk was studied by using visible/near-infrared (VIS/NIR) reflectance spectroscopy (350～2 500 nm). The spectral data and the protein content data of the pure milk samples were collected by ASD spectrometer and CEM rapid protein analyzer, respectively. Based on the analysis and comparison of different spectrum preprocessing methods and band selection methods, the feature bands were determined. Finally, using the Principle Component Regression (PCR) and Least Squares Support Vector Machine (LS-SVM) model, the regression models between the reflectance spectroscopy and the protein content in milk were presented for pure milk samples and the predictive ability was also analyzed. In this way, the optimal inversion model for true protein content in milk was established. The results were shown as follows: (1) In the process of spectral pretreatment, the combination of multiple scatter correction and second derivative achieved a better result; (2) Compared with the modeling of whole spectral, appropriate variable optimization models had the ability to improve the accuracy of the inversion results and reduce the modeling time; (3) The analysis results between PCR model and LS-SVM model demonstrated that the prediction accuracy of LS-SVM model was better than PCR model. The coefficient of determination (R2P) of PCR and LS-SVM were 0.952 2 and 0.958 0 respectively, and the root mean square error of prediction (RMSEP) of PCR and LS-SVM were 0.048 7 and 0.048 2 respectively. The result of this research is expected to provide a novel method for nondestructive and rapid detection of true protein in milk.