Federated learning (FL) models face challenges such as negative transfer in non-IID scenarios, where the global model performs worse than local models. Clustered Federated Learning (CFL) addresses this by grouping clients based on pairwise similarity, but the high computational cost limits the scalability. To solve this, we propose FedBCE, an efficient CFL framework that significantly reduces clustering time by calculating distances only to selected centers while grouping clients using data volume and model similarity to enable collaboration among those with similar data distributions. Experiments on public data sets show that FedBCE reduces the clustering time to just \(\frac{1}{10}\)  to \(\frac{1}{160}\) of existing methods while increasing the accuracy by 19% on average in non-IID scenarios.