This paper presents the use of Long Short-Term Memory (LSTM) networks as models for prediction in the Model Predictive Control (MPC) algorithm. LSTMs are recurrent neural networks often used to model dynamical processes. MPC is an advanced control technique using a process model to calculate online predictions. MPC sets the control policy for the process by solving an optimisation problem to minimise the prediction error and not violate the constraints. This paper compares two LSTM model architectures for MPC in a process with Multiple Inputs and Multiple Outputs (MIMO): a single LSTM model with multiple inputs and multiple outputs (LSTM MIMO) and several parallel LSTM models, each with Multiple Inputs and Single Output (LSTM MISO). The quality of modelling using these two architectures is analysed. Next, the selected MIMO and MISO LSTM models are implemented in the MPC algorithm. Finally, the control quality and execution time are investigated. It is concluded that both MIMO and MISO approaches offer distinct advantages; however, for the benchmark neutralisation reactor, MIMO models provide a more efficient solution for MPC implementation.