Hybrid system identification aims at both estimating the discrete state or mode for each data point, and the submodel governing the dynamics of the continuous state for each mode. The paper proposes a new method based on kernel regression and Support Vector Machines (SVM) to tackle this problem. The resulting algorithm is able to compute both the discrete state and the submodels in a single step, independently of the discrete state sequence that generated the data. In addition to previous works, nonlinear submodels are also considered, thus extending the class of systems on which the method can be applied from PieceWise Affine (PWA) and switched linear to PieceWise Smooth (PWS) and switched nonlinear systems with unknown nonlinearities. Piecewise systems with nonlinear boundaries between the modes are also considered with some preliminary results on this issue.