Restricted by the data sensing capability, it is challenging for a single user to generate high-quality deep learning models based on its collected few-shot data samples. Meta-learning provides a promising paradigm to make full use of historical data at the base stations to improve the performance of few-shot learning tasks. However, it is a dilemma to balance the performance and the communication costs of meta-learning. In this paper, we studied the design of few-shot learning in wireless networks. First, a meta-learning model-based scheme is designed to adapt the few-shot learning tasks, and a multicasting-based model transmission scheme is proposed. Second, a coalition formation-based model selection scheme is designed to achieve a sophisticated tradeoff between the performance and the communication costs of meta-learning. Finally, the simulation results are provided, which show that our proposed scheme can improve the model accuracy performance with low communication costs.