In this paper the problem of order selection for nonlinear dynamical Takagi-Sugeno (TS) fuzzy models is investigated. The problem is solved by formulating the TS model in its Linear Parameter Varying (LPV) form and applying a recently proposed Regularized Least Squares SupportVector Machine (R-LSSVM) technique for LPV models. In contrast to parametric identification approaches, this non-parametric method enables the selection of the model order without specifying the scheduling dependencies of the model coefficients. Once the correct model order is found, a parametric TS model can be re-estimated by standard methods. Different re-estimation approaches are proposed. The approaches are illustrated in a numerical example.