Traditional POI recommendation systems rely on centralized data storage, raising privacy concerns. Federated learning addresses these issues but assumes clients have sufficient resources, neglecting limitations of mobile devices like smartphones, which causes delays, storage shortages, and high communication overhead. This paper introduces LFR-POI, a lightweight federated framework. It uses a personalized discrete module to encode latent features as binary vectors for efficient computation in Hamming space, and a client clustering method based on Hamming similarity to address data heterogeneity. In addition, pseudo-labeling techniques are adopted to protect privacy. Experimental findings show that LFR-POI achieves better performance than other federated baseline approaches on three datasets.