A beef quality on-line detection and classification models by Vis/NIR reflectance spectroscopy was established. The system parameters were optimized. Signal collection and spectroscopy preprocess were carried out by Vis/NIR reflectance spectroscopy and a handheld probe device. The scanning times were set on condition that system kept proper detection accuracy and stability, which was 10 times in wavelength range of 400~700nm and 30 times in wavelength range of 700~2000nm, and acquisition time of 900ms. Spectra leverage value of beef was calculated to eliminate abnormal samples, and then different data processing methods were used to establish beef quality PLSR models which finally showed the optimal result of beef quality prediction. The results indicated that the PLSR model with SNV processing had better performance, with the correlation coefficient of 0.9068 and root mean square error of 7.1963N for validation set of beef tenderness, 0.8854 and 2.3628 for L*, 0.8362 and 2.2969 for a*, 0.8453 and 2.1054% for validation set of beef cooking loss, respectively. The correlation coefficient was above 0.8 and the tenderness classification accuracy reached to 93.5%.