This paper examines how elastic model sizes can accommodate devices with different hardware resources in federated learning.