The invention discloses a music model training method. The music model training method comprises the following steps: acquiring a MIDI music data set, wherein the MIDI music data set comprises a plurality of MIDI music scores; extracting the feature vector of each MIDI music score; inputting the feature vectors into a structured support vector machine for training, so that a music model is obtained, the step specifically comprises the following substeps: constructing a discrimination function f(x;w), wherein x is a feature vector, w is a parameter vector, carrying out outputting by adopting the data value (with the calculation formula shown in the description) of the maximal discrimination function f(x;w) as the predicted value, calculating the predicted value and the true value accordingto a preset loss function (shown in the description), wherein P is the probability distribution of data, which is replaced with the empirical risk (shown in the description) obtained through calculation with the trained sample data, solving the unique parameter vector omega by adopting the optimizing formula (as shown in the description) of SVM, so that the empirical risk (shown in the description) obtained through the trained sample data is 0, solving the discrimination function f(x; omega), and finally, outputting the music time sequence. The invention further provides a music creation method, devices, a terminal and a storage medium. In the technical scheme, artificial intelligence is used for music model training for the first time, for the trained music model, the feature extraction capacity of the MIDI music score can be improved.