Existing works on vehicular edge computing (VEC) fail to co-consider selfishness of vehicles and characteristic of deep neural network inference, which results in a lack of guarantee in quality of inference services. This paper proposes an efficient incentive algorithm for model deployment and task offloading in VEC to improve quality of inference services. Specifically, we firstly formulate a problem with the objective of maximizing the social welfare under constraints of accuracy requirement, latency requirement, etc. Then, we propose a coalition formulation-based auction algorithm, which is based on coalition game and Vickrey-based auction mechanism, to solve the formulated problem. Simulation results show that, the social welfare of the proposed algorithm can be successfully increased by \(19.90\%\) and \(33.89\%\) on average for different numbers of RSUs, compared with the two state-of-the-arts.