In this work, we study how to keep federated learning useful when every participant has a different link, a different device and a different rule book. Classical solutions still sit at two extremes. Centralised schemes count on one server, which quickly becomes a bottleneck and a single point of failure, while fully decentralised schemes slow down as soon as the graph turns sparse. We introduce heterogeneous communication in Decentralised Federated Learning in a model named Hybrid Hierarchical Decentralized Federated Learning, a hybrid hierarchical protocol that splits the traffic into two cooperative layers. Locally, each client shares only one chosen layer with a lightweight coordinator inside its organisation; the coordinator stitches the layers together and sends back a full model. Globally, those coordinators swap complete models with their peers through a push-sum routine, so the system stays fully decentralised. This split lets us tune the size and cadence of the messages at each layer, keeping client traffic modest and protecting organisational autonomy while still letting knowledge flow across the network. A concise convergence sketch suggests that the error gap with respect to FedAvg remains bounded even under non-IID data. We believe this design offers a practical route for federated learning in real scenarios such as hospitals, campuses or edge swarms where bandwidth, trust and regulation differ from place to place.