The viscosity of the protective slag has a direct influence on the surface and internal quality of the cast billet. To precisely predict the viscosity value of the protective slag and assist in the composition design of the protective slag, a viscosity prediction algorithm based on a gated recurrent unit (GRU) model with soft attention was put forward. Firstly, a dataset of fluoride-free protective slag composition was collected, and the Pearson correlation analysis method was employed to conduct a correlation analysis between each component in the protective slag and the viscosity value. Then, the soft attention mechanism was integrated into the input processing stage of the GRU model to enhance the input features and improve the capture of global feature information. Finally, viscosity prediction experiments were carried out using the fluoride-free protective slag dataset, and the optimized algorithm was compared with the classical prediction algorithms to verify the validity of the GRU model integrated with the soft attention mechanism. The experimental results demonstrate that under the same hyperparameters, the average absolute percentage error of this model is 7.78%, the coefficient of determination is 0.906 5, and the mean square error is 0.003 3, all of which are higher than those of the comparison models. The prediction accuracy and stability have been significantly enhanced, making it more applicable for the prediction of the viscosity of fluoride-free protective slag.