In order to realize fast, nondestructive and real-time detection of nutrition components (fat and protein) for pork, a portable nondestructive detection device based on near infrared reflectance spectra was designed and developed. The hardware part included spectrum acquisition unit, light source unit and control unit. The corresponding detection software was developed to realize the effective acquisition and real-time analysis of the sample spectrum information. In order to establish a stable and reliable forecasting model, the research focused on the effects of band selection, different sample grouping methods and variables selection methods on the models. Based on visible/short wavelength near infrared (Vis/SWNIR), long wavelength near-infrared (LWNIR) and Vis/SWNIR-LWNIR, all the samples were divided by random selection (RS) method, Kennard-Stone (KS) algorithm and sample set partitioning based on joint X-Y distances (SPXY) algorithm, and then partial least square prediction models for fat and protein content were built, respectively. The results showed that the best prediction models for fat and protein were built based on Vis/SWNIR-LWNIR by using SPXY algorithm. On the basis of the best model for each parameter, comparative analysis of competitive adaptive weighted algorithm, Random Frog algorithm and uninformative variable elimination-successive projection algorithm were employed to screen variables. The results showed that the simplified model based on competitive adaptive weighting algorithm was the best with correlation coefficients in the prediction set of 0.9505 and 0.9510 for fat and protein, respectively. The results indicated that the designed portable detection device based on near infrared reflectance spectroscopy was able to realize fast, nondestructive and real-time detection of fat and protein content for fresh meat and had certain application potential and market prospects.