While data-driven fluid simulation methods greatly replace the physics-based fluid solver and achieve high quality results, it is a challenge to get enough realistic effect with less time. The Huge neural network models brought by the complexity of fluid data need to calculate a large number of parameters from the convolutional and full-connected layers in the forward propagation process, which lead to very long inference time and cannot meet the real-time requirements. Our method is based on a structural pruning method to reduce the number of parameters of a general fluid neural network model that imposes the admm constraints on original loss on training process and removes the convlutional filters at a certain rate according to their importance. We show the high quality results for velocity field reconstruction and advancing time from reduced parameters using our pruned fluid model, which has only 30%–50% parameters of the original model and greatly improves the inference speed of the model. It is a big step towards high-accuracy real-time fluid simulation.