Federated learning has emerged as a promising paradigm to train models collaboratively without data centralization. However, tackling both non-IID data distributions and diverse model architectures remains a formidable challenge for FL. To address these challenges, existing methods have demonstrated promising performance by facilitating knowledge sharing through class representations (e.g., prototypes). Nonetheless, these approaches often rely on a single global prototype, which fails to capture localized nuances across clients, thereby leading to ineffective knowledge transfer. To tackle these limitations, this paper proposes a novel similarity-based prototype aggregation mechanism (FedSPA), which adaptively generates multiple personalized prototypes by evaluating cross-client similarity, thereby distilling relevant knowledge while preserving local adaptations in diverse networks. Through extensive experiments, we demonstrate that FedSPA achieves significant performance improvements, with accuracy gains of up to 8.74% and a 41.07% reduction of communication overhead.