The device used in the present study consists of a conveyer belt that throws the fruit onto a flat horizontal plate connected to a load cell. The vertical distance between plate and conveyer belt (drop height) as well as the speed of the belt can be continuously adjusted. Tests were carried out by selecting three different values of drop height and speed. The Magness-Taylor (MTf) index was used as reference, destructive parameter, to describe the flesh firmness and to set-up predictive models. The digitalized time history of the force was analysed to extract some mechanical indices (peak force, impact duration and impulse) used to predict MTf by simple or multiple regression analyses. Moreover, each point of the entire time history was processed by artificial neural network (ANN) software to predict MTf. The goodness of fit, expressed as R(2), was up to 0.823 with the regression models. On the whole, the peak force was the best predictor. The ANNs did not involve a substantial increase in goodness of fit with respect to the best regression models: +83%, as mean, 37% as maximum. The speed or position at which the fruit impacts the plate can represent an important parameter influencing the MTf prediction. Free dropping of the fruit instead of throwing onto the plate by the conveyer did not provide a better prediction. The impact device did not cause mechanical damage to the kiwifruits. (C) 2009 Elsevier Ltd. All rights reserved.