A multiscale v-support vector machine(v-SVM) based on adaptive boundary vector extraction is presented. It overcomes the disadvantages of the slow training speed and the low regression accuracy which are caused by using the general v-SVM for large-scale and multi-peak sample modeling. An adaptive boundary vector extraction algorithm is used to extract the boundary vectors which include all support vectors from the training samples, so that reduces the sample scale. The global optimal regression model is obtained by solving the multiscale v-SVM quadratic programming problems, and the complex distribution sample can be approximated from multiple scales by the model. Simulation results show that the v-support vector machine based on boundary vector extraction has better regression results than the general v-SVM.