The behavior of Model Predictive Control (MPC) is by principle affected by the quality of the model used for the controlled system. A parameter mismatch between the plant and the controller can affect drastically the performance of MPC. This paper presents a new strategy called model-free predictive control (MF-PC) to overcome these problems. In this approach, a recursive least squares algorithm is implemented to identify the parameters of an auto-regressive with exogenous input (ARX) model, using input and output measurements of the controlled system. The proposed method enables an accurate prediction of the controlled variables without requiring detailed knowledge about the physical system. Simulation results obtained for the current control of a two-level, three-phase voltage source inverter, demonstrate that the MF-PC is exceptionally robust against the parameter variations and the model uncertainties, compared to conventional finite-control-set MPC.