The predictive ability of the partial least squares（PLS） model established by near-infrared spectroscopy is often limited when there are changes in the composition of the test sample or environmental conditions，making it difficult to predict new samples accurately. In this case，the labeled samples containing the new changes need to be added to the calibration set for model update. However，the large size of the old calibration set combined with the limited number of new samples can make it difficult for the new changes to be reflected in the model. To be able to update the model quickly，this paper proposes a method to reduce the number of samples by selecting important old samples using kernel coefficients（KCS）. A kernel model is built for the old samples to obtain the coefficients of each sample in the model，larger coefficients correspond to greater sample importance. Thus，samples with higher coefficients are chosen to reduce the sample size，and new samples are added to the part of old samples to update the model. The experiment compared the model updates of KCS by including a portion of old samples with the model updates by including all old samples with the new samples. The experiments were conducted on simulated and soybean meal datasets. Using a portion of old samples for model updates in KCS decreased the root mean square errors（RMSE） of the predictions from 1.165 and 0.730 before the update to 0.961 and 0.654 after the update，which represents a decrease of 17.5% and 10.4%，respectively. Using all old samples for model updates in KCS decreased the RMSE of the predictions from 1.110 and 0.720 before the update to 0.980 and 0.662 after the update，which represents a decrease of 11.7% and 8.1%，respectively.The results show that this method of selecting some important old samples for model updating solves the imbalance of the number of new and old samples，thereby accelerating the update speed of models.