Measurement speed is a key fact that limited the practical application of RF device behavior model. In this paper, a novel rapid extraction method based on reverse wave injection is proposed. This method changes the amplitude and phase of the reverse incident wave at the load side of the device under test, so that massive training data can be acquired for the Bayesian algorithm, without the need for a complete load-pull test. Experiment is conducted over a GaN device at 2GHz frequency, the results prove that the method has greatly increase the extracting speed without compromising the accuracy. Compared with the traditional scheme, this method can greatly improve the model extracting efficiency and can be widely applied in practice.