Federated learning (FL) enables effective model training across distributed devices while protecting data privacy. However, high communication costs challenge practical FL, particularly with computing-capability-limited local devices. To address this, we propose a novel FL strategy in device-to-device (D2D) networks with model pruning. Dynamic clustering using k-means algorithm enhances information exchange efficiency, while model pruning reduces local model complexity and training latency. We derive a convergence bound of $\mathcal{O}\left( {{t^{ - 1}}} \right)$ and optimize the pruning ratio and bandwidth allocation using KKT conditions. Simulation results reveal that without the sacrifice of testing accuracy, the proposed FL algorithm requires only 88% of the training time compared to the system employing equal pruning, and 80% of the training time compared to the system without pruning.