This research proposes a communication-efficient federated learning (FL) framework, leveraging adaptive model aggregation to optimize the balance between communication cost and performance of the model. The methodology employs dynamic client selection and aggregation strategies to reduce unnecessary communication while maintaining model accuracy. Clients train local models on decentralized data; only significant updates are transmitted to the central server for aggregation. By adjusting communication intervals and evaluating the quality of client updates, the framework minimizes overhead without compromising model convergence. Experimental results demonstrate a significant reduction in communication rounds, achieving high model accuracy (over 90%) with fewer updates. The framework is scalable and suitable for real-world applications with constrained bandwidth and computational resources. Overall, this approach enhances the efficiency of federated learning by dynamically adjusting communication strategies, ensuring high performance and cost-effectiveness.