Federated learning (FL) has become a promising technique for developing intelligent wireless networks. In traditional FL paradigms, local models are usually required to be homogeneous for aggregation. However, due to heterogeneous models coming with wireless sysTem heterogeneity, it is preferable for user equipments (UEs) to undertake appropriate amount of computing and/or data transmission work based on sysTem constraints. Meanwhile, considerable communication costs are incurred by model training, when a large number of UEs participate in FL and/or the transmitted models are large. Therefore, resource-efficient training schemes for heterogeneous models are essential for enabling FL-based intelligent wireless networks. In this paper, we propose an adaptive quantization scheme based on ensemble distillation (AQeD), to facilitate heterogeneous model training. We first partition and group the participating UEs into clusters, where the local models in specific clusters are homogeneous with different quantization levels. Then we propose an augmented loss function by jointly considering ensemble distillation loss, quantization levels and wireless resources constraints. In AQeD, model aggregations are performed at two levels: model aggregation for individual clusters and distillation loss aggregation for cluster ensembles. Numerical results show that the AQeD scheme can significantly reduce communication costs and training time in comparison with some state-of-the-art solutions.