Abstract: The embodiments of the present disclosure provide a method and apparatus for training an SVM model, a device and a computer-readable storage medium. The method comprises steps of: acquiring a dataset; loading the dataset; predicting the best storage format of the dataset online through an automatic tuning framework, and performing format conversion; and, executing an SMO algorithm to train an SVM model. In this way, the best data storage format and algorithm can be automatically selected according to the input dataset; in a global loop level, a large number of more efficient matrix multiplication operations are performed on the basis of the best data storage format to calculate a kernel matrix; and, in a local SMO solver level, finer memory optimization is performed by the register communication technology and the combined memory access method of the GPU, and the calculation resources of the hardware platform are fully utilized.