Grinding circuit needs to be stably controlled for high recovery rate of mineral ore and significant reduction of production cost in concentrator plants. Ball mill grinding circuit is essentially a multi-input–multi-output (MIMO) system with strong coupling among process variables. Simplified model with multi-loop decoupled PID control usually cannot maintain a long-time stable control in real practice. The response tests between four controlled variables (namely, product particle size, mill solids concentration, sump level and circulating load) and four manipulated variables (namely, fresh ore feed rate, mill feed water flow rate, pump speed and dilution water flow rate) are carried out to construct a four-input–four-output model of grinding circuit. Based on this modeling, constrained model predictive control (MPC) is adopted to handle such strong coupling system and evaluated in an iron ore concentrator plant. The variables are controlled around their set-points and a long-term stable operation of the grinding circuit close to their optimum operating conditions is achieved. More than three years’ operation in industry demonstrates the effectiveness and practicality of this control strategy.