Nonlinear friction caused by the flexible link and the transmission process in the harmonic drive leads to complex hysteresis characteristics of harmonic transmission that inevitably affect the transmission accuracy. To describe the hysteresis characteristics of the harmonic drive, a concise neural network hysteresis hybrid model, comprising hysteresis-like characteristic preconditioning in series with a dynamic neural network, was presented in this study. It was executed in two steps: the input signal was preprocessed to produce hysteresis-like behavior; the dynamic Radial Basis Function (RBF) neural network was fully utilized to achieve high-precision approximation of hysteresis-like to hysteresis characteristics of the harmonic drive. Moreover, an experimental platform was constructed in this study, and the data obtained under different experimental conditions were modeled and verified. Both at a constant input accuracy and the accuracy with different input signals and loads, the verification accuracy obtained by the neural network hysteresis hybrid model is 0.449 6 (Mean Square Error (MSE)), which is much higher than the 3.0321 (MSE) accuracy of the classical neural network model. This proves the effectiveness and adaptability of the proposed neural network hysteresis hybrid model.