Model predictive control (MPC) is an online application based on dynamic models. Its application faces two major obstacles: (i) computational constraints and (ii) the need to accurately simulate the process by a model that properly predicts how the plant will behave in the future. Implementation of MPC is not always possible in large-scale or industrial applications due to the computational complexity of MPC and to the dimensionality of the models. To facilitate MPC implementations, this paper proposes a self-adaptive approach based on simplified (or reduced-order) nonlinear models. The proposed methodology yields an MPC that adjusts the dimension of the model according to both the current process conditions and the control objectives. The self-adaptive approach is described and validated on an industrial case study, a C4-splitter.