Abstract: The optimization process of automobile aerodynamic design that requires a vast amount of CFD data, which are both expensive and time-consuming to acquire, is arduous and costly. To get out of this predicament—the conflict between accuracy and efficiency—we devise an innovative automobile shape optimization technique relying on the Multi-fidelity deep neural network (MFDNN) aerodynamic model based on transfer learning and data fusion. Applying the developed optimization method to the drag reduction optimization design of the fast-back MIRA standard model demonstrates that the method can fully integrate the knowledge contained in different fidelity data, accelerate the aerodynamic shape optimization process, and improve the optimization efficiency. Specifically, the convergence speed of the optimization framework based on multi-precision neural network is 5.85 times that of the offline optimization framework based on single-precision neural network and 2.81 times that of the online optimization framework based on single-precision neural network.