Abstract: A direct way for getting controller0s coe±cents of multivariable generalized predictive control (MGPC) of diagonal CARIMA model is developed in order to simplify the design and implementation of MGPC. A very concise MGPC controller is obtained by directly manipulating the model predictor of a diagonal multivariable CARIMA model recursively, and the control moves are the product of the controller0s coeffcients and set-points, historical input/output data of the plant and predictive errors of the predictor. The controller's coeffcients are determined only by the model parameters and design parameters, and the number of coeffcients only depends on the orders of the model. This method avoids solving Diophantine equations, reduces the computational overhead on line, and simplifies the implementation of MGPC. Its validity is demonstrated by comparative experiment results obtained from a nonlinear liquid system controlled by DCS.