Nowadays, Federated Learning (FL) is widely applied in the Internet of Things (IoT). However, when a large number of devices participate in FL, they still face the challenge of low communication efficiency. In addition, how to reasonably allocate FL trained models to third parties (e.g. task publishers) is also a problem that needs to be solved. In this article, we propose a hierarchical FL (HFL) framework based on incentive mechanisms, where task publishers mobilize users for collaborative computing through edge servers. At the lower layer, evolutionary game is used to model the dynamic decision-making process of users with bounded rationality, and users select user groups (UGs) to participate in training by considering model accuracy and training costs. At the upper layer, an iterative double auction mechanism is adopted to allocate the model reasonably to multiple task publishers, maximizing the total social welfare. Finally, the effectiveness of the proposed scheme is verified through experiments.