Nowadays, the sorting of agricultural products based on their quality indices is very important. Kiwifruit is one of the valuable fruits in Iran and its production is considerable compared to the other countries. The non-existence of a standard fruit sorting and packing systems is the main problem in post harvest technology and exportation of this product. Among different quality indices, weight is one of the most important ones which its rapid estimation is of great interest in sorting lines as well as estimation of other quality indices such as volume and maturity. Several researches have been reported in developing a high speed weighing systems of agricultural product. In the most of them, they have used the non-destructive impact method using accurate and sensitive load cells for this purpose. But their configurations and modeling approaches that were used may be quite different. In this research, the ability of a conveyor- load cell unit based on falling impact method was evaluated in rapid weighing of kiwifruit. Sufficient amount of fruits (232 samples) with a versatile range of weights (40 -120 gr) were selected during harvest time. The samples were then weighed with a digital balance with an accuracy of 0.001 gr. Immediately, the impact signal of the samples were acquired with the mentioned system. All the test were carried out at three different forward speeds (1, 1.5 and 2 m/s) of conveyor belt. Three different modeling methods, i.e., multiple regression, partial least squares (PLS) and combined principal components analysis (PCA) with artificial neural networks (ANN) were used in extracting the weight predictor models and their results were compared. The multiple regression analysis were carried out using the impact indices (peak force, F p , peak duration D p , and peak impulse, I p ) of the first peak as well as the impact indices of the first forty successive peaks of the impact signal. The PLS analysis were carried out using the entire signal data points and also the first peak information as the input variables. The results showed that impact indices of the first peak did not result to an acceptable prediction powers while by using the forty successive peaks, the prediction power improved considerably to a R 2 of 0.860 and a SDR of 2.857 at 2 m/s forward speed. The results of PLS regression with entire impact signal showed the increase in prediction power with a R 2 of 0.936 and a SDR of 3.865 at 1.5 m/s conveyor speed. The PCA-ANN regression analysis also resulted to a slightly better results compared with PLS analysis at 1 m/s forward speed with R 2 of 0.915 and SDR of 3.451. As a consequence, the PLS and PCA-ANN models led to the best weight prediction models while the multiple regression with impulse values of first forty peaks, PLS model with first peak information and single regression with impact indices of the first peak had the next positions in weight prediction models.