Meat samples for calibration (n=210, pork, beef and mutton n=70, respectively) were scanned, over a NIR spectral range of 4 000～10 000 cm^-1, and the chemical analysis were performed. Meat samples (n=90, pork, beef and mutton n=30 respectively) were scanned and analyzed for prediction of protein content. It was developed a PLS regression model assaying based on different spectral pretreatment methods. The best calibrations models of fresh meat samples showed relatively good predictability for protein, the coefficient of determination of calibrations samples was 0.954, the coefficient of determination of prediction samples was 0.929, the RMSEC and RMSEP were 0.495 and 0.669, respectively. Therefore, the fresh meat quantitative models can apply for protein prediction for different meat samples, which enhanced its application range.