A new Implicit Generalized Predictive Control (IGPC) algorithm based on online Least Square Support Vector Machines (LSSVM) inverse control is proposed. Firstly, an offline model of original nonlinear system is obtained. Online LSSVM is used to identify αth-order inverse dynamic model of nonlinear systems, which can compensate the errors of nonlinear systems caused by offline identification adaptively. Then the model of online αth-order inverse plant is cascaded before positive plant to create an αth-order delay pseudo-linear composite system, which can complete decoupling and linearization of multivariable systems. Then an implicit generalized predictive control is used to control the pseudo-linear composite system. Inputs are constrained in the whole of prediction horizon and control horizon. The simulation and experiment results for the typical nonlinear system and supercritical 600MW CFB process are shown that IGPC of multivariable systems based on online LSSVM have better tracking and strong anti-interference performance.