Prediction ability of visible/near-infrared spectroscopy for intramuscular fat content assessment on-line was studied. Totally 208 longissimus dorsi muscle samples were collected from different carcasses. Sample spectra were scanned while the samples were moving at the speed of 0.25m/s. Wavelet transform was employed to eliminate the spectra noise. Partial least squares regression (PLSR) based on the de-noised spectra, combined with different pretreatment methods, was explored to predict intramuscular fat. Daubechies of vanishing moment6 at decomposition level 6 with the minimax threshold gave good de-noising results. PLSR model based on the de-noised spectra without any pretreatment had poor prediction performances. And through pretreatment, including multiplicative signal correction (MSC), standard normalized variate (SNV) and derivative, the performances are improved. The first derivative spectral in conjunction with SNV yielded the best PLSR model with correlation coefficient of 0.892 and 0.834 in calibration and validation sets, respectively, alongwith root mean square error of calibration of 0.090 and root mean square error of 0.080 for validation. Results indicate that it is possible to predict intramuscular fat content on-line. However, the best ratio of the standard deviation to the SEP of the validation set (RPD) value was 1.738, and the accuracy and the robustness of the model needed further improvement for practical on-line application.