Abstract: The work aimed to improve the prediction efficiency in rapid determination of the firmness of kiwifruit. A Vis/NIR (390~1030 nm) hyperspectral imaging system was applied to obtain the images of Guichang kiwifruit, and the reflective spectrum in the regions of interest on each sample was acquired. Noise from original reflective spectrum was reduced by the standard normal variation method. The competitive adaptive reweighted sampling (CARS) and the successive projection algorithm were applied to select feature variables. Finally an error back propagation neural network and a multi linear regression (MLR) model were constructed to predict the firmness of kiwifruit. A total of 35 feature variables were selected by CARS from 256 variables. The working efficiency of the final prediction model was improved by 11-fold, with the runtime dropped from 5.84 s to 0.54 s. Overall, the CARS-MLR model showed a relatively good detection capability (r_c=0.95, r_p=0.92, RMSEC=1.65 kg/cm², RMSEP=1.99 kg/cm², RPD above 2). This study demonstrated the application potential of the nondestructive hyperspectral imaging technology for fast determination of the firmness of kiwifruit.