A nonlinear predictive generalized minimum variance (NPGMV) control algorithm is introduced for the control of nonlinear multivariable systems. The plant model is represented by a series combination of a nonlinear operator, which is assumed finite-gain stable, and a linear state-space model, which can include time delays and unstable modes. The main contribution is to incorporate predictive action into the recently introduced Nonlinear GMV controller by defining a multi-step cost index and using a minimum-variance form of the usual GPC cost function. The solution is very different to traditional nonlinear model predictive control, providing a solution which is similar to fixed model based controllers. This does not provide the same constrained optimization features but it does give a controller which is very simple to implement.