Federated Learning (FL) is a distributed machine learning method with great potential for the Vehicular Networks. However, traditional FL involves frequent exchanges between RSUs and vehicles, causing high communication overhead. Additionally, the Non-IID nature of vehicle data, due to varying driving habits and environments, reduces model accuracy. To address these issues, this paper proposes a Semi-Asynchronous Federated Dynamic Mutual Distillation (SFMD) method. It first selects the top N vehicles for model updates based on processing time and forces inactive vehicles to join training. Secondly, teacher and student models perform mutual knowledge distillation with dynamic intensity to improve generalization and training. Finally, a data sharing mechanism based on differential privacy reduces model bias from Non-IID data. Simulations show SFMD improves model accuracy and reduces communication overhead.