Federated Learning is one of the main research lines in the last years about distributed learning, where participating nodes share their models but maintain the privacy of the data used to learn such models. Consensus is a way of calculating a mean value between a set of agents using only information from the local neighbors. This paper presents a new approach based on Asynchronous Consensus, called Multi-layered Asynchronous Consensus-based Federated Learning (MACoFL). It randomly chooses a neighbor and a layer from the neural model and interchanges it with him. This new algorithm is presented and tested using the MNIST dataset.