Abstract: Intelligence and distribution are the main directions of future warfare, while intelligent model is an important means to realize them. In the battlefield, enemy operations have brought huge challenges to the evolution of intelligent models. Therefore, an architecture for continuous evolution of intelligent models for distributed operations is proposed. This architecture can adaptively select the centralized hierarchical federated learning and the fully distributed gossip learning according to the communication status. And dynamically switch as the battlefield environment changes. It has been verified that this architecture can adapt to different communication conditions and achieve a better intelligent model evolution effectiveness.