Federated learning enables the training of machine learning models in a multiple-client single-server setting by exchanging model weights or gradients over the network, thereby preserving data privacy. However, this paradigm introduces significant communication overhead, which represents a major bottleneck in practical deployments. Existing approaches, such as pruning and quantization, aim to mitigate this cost, while recent research has explored the use of machine learning models themselves to regenerate model parameters, notably through hypernetworks. In this work, we propose a federated learning architecture that combines client-side representation summarization with server-side weight regeneration. Specifically, an IJEPA model is employed on the client to extract and summarize information from the embedding space, while a hypernetwork on the server reconstructs model weights based on the received information.