The application discloses a method and device for updating a deep learning model, and a system. One specific implementation mode of the method comprises the following steps: a new training data set sent by a user terminal is received, wherein the new training data set is detected by the user terminal in a preset path; a preset deep learning model is trained according to the new training data set to obtain a prediction model after training; and the preset deep learning model is updated to the prediction model so that the prediction model is used in online execution data prediction operation. Docking with the training data set of a user is achieved through the implementation mode, and updating efficiency of the deep learning model is improved.