With the emergence of the sixth generation (6G) communication technologies, massive infrastructures will be densely deployed and the number of data will be generated exponentially. Massive data collected by vehicles can be used to train a machine learning model, which is an effective way of implementing intelligent services in the Internet of Vehicles (IoV). However, centralized training leads to communication congestion and user privacy leakage. To address these problems and improve efficiency, federated learning has been proposed. Most current studies on federated learning consider all users performing only a single task at a certain time, which results in low network utility and long task execution time. Because some vehicles may have similar data, we propose a multi-task federated learning model where vehicles simultaneously execute tasks. To jointly maximize the network utility and efficiency of the multi-task federated learning in the 6G-enabled IoV, we design a task-driven vehicular coalition algorithm considering vehicle selection and wireless resource allocation. Moreover, we derive a convex objective function from network utility function and loss function with constraints through a series of mathematical analyses. Finally, we verify that our proposed algorithm with low complexity can improve the utility and efficiency of the multi-task federated learning in the 6G-enabled IoV.