To predict and classify beef tenderness, a laboratory hyperspectral imaging system was developed to capture hyperspectral scattering images from the surface of beef samples in the spectral region of 400～1000nm. Reflectance spectral characters were obtained from hyperspectral image. By using the method of step-wise regression, six optimal bands, 430, 496, 510, 725, 760 and 828 nm were selected for establishing the multi-linear regression (MLR) model. The model gives good prediction values of beef WBSF with the correlation coefficient of cross validation of 0.96 and the standard error of cross validation of 0.64 kg. Based on the measured tenderness values, samples were divided into two classes, i.e., group 0 (<6.0 kg) and group 1 (>6.0 kg). From selected bands, canonical discriminant functions were built to divide samples into two classes. The full cross validation was employed with the classification accuracy of 83.3% and 90.9%. Resultingly, the overall accuracy of classification is 87.0%. This research demonstrates that the hyperspectral imaging technique is useful for nondestructive determination of beef tenderness.