The raw spectra of milk powder were detected by means of two Fourier near-infrared spectrometers MPA type and WQF-400N type. Primitive spectra were integrated in different integral intervals to simulate spectra with different resolutions. After being pre-treated with the same methods (Savitzky-Golay smoothing and the first derivative),models for fat content in milk powder were established through partial least square regression (PLS). Comparisons on correlation coefficient R,root mean square error of prediction fRMSEP and mean bias E between the models reveal that MPA type and WQF-400N type spectrometers have the best models at the resolutions of 16 cm-1 and 64 cm-1 respectively. These results show that integral spectrum could compress data and enhance the model′s prediction accuracy,and that an adequate resolution should be selected for special characters of different samples. Both of them ensure the best quantitative analytical result while reducing the work load. It is of no significance to simply pursue higher and higher resolution.