Portable near infrared spectrometer was applied to collect 320 pieces of fresh pork spectral information in NIR medium wavelength region. Prediction models of fresh pork cholesterol level with NIR spectroscopy were established through partial least squares method combined with different spectroscopy preprocessing methods. The effects of outlier samples elimination and combination of different preprocessing methods on the prediction model performance were discussed. The result showed that the optimum prediction model of fresh pork cholesterol level was achieved with the application of two optimization procedures, eliminating outliers twice and combination of SG first order derivative, SG smoothing and orthogonal signal correction, and the relevant parameters as follows: Rc=0913 7, SEC=2560 7, Rp=0656 7, SEP=4985 5, MF=4, RPD=2503 2, RSEP=8625 4%, SEP/SEC=1946 8, which indicated the reliability, resolution capacity and prediction accuracy of this model in NIR medium wavelength region were acceptable. The robustness of optimal prediction model could be further improved by adding more representative and typical sample of different cholesterol level range into the calibration set. Paired-samples t-test showed non-significance between the predicted value and reference value (p＞005), and the total prediction accuracy of testing samples was 625%, and partial prediction accuracy was 917% in cholesterol range of 50～70 mg·(100 g)-1, which showed that this model could be applied into on-line rapid preliminary quantitative analysis of cholesterol level of fresh pork. In this research it was the first time that portable near-infrared spectrometer was applied into the analysis and detection of cholesterol level of fresh pork products within NIR medium wavelength region, and with further study and improvement, the prediction model could also be applied to raw material classification, quality and process control, random inspection of commercially available meat and meat products.