This paper proposes a navigation method for unmanned aerial vehicles (UAVs) that integrates multimodal large model reasoning with autonomous exploration technique for tackling zero-shot target searching in unknown environments. First, to address the difficulty of multimodal large models in directly processing 3-D data, a space–vision inverse mapping approach is introduced. By constructing scene images with explicit 3D coordinate constraints as inputs, the multimodal large model is endowed with the capability to simultaneously understand scene imagery and localize key regions. Second, to overcome the poor generalization of existing target searching methods, a prompting strategy embedding a “recognition–evaluation–transfer” logic is designed to guide the UAV in performing zero-shot target searching across diverse scenarios. Finally, to mitigate the significant sim-to-real gap that existing methods are currently struggling with, a geometric–semantic asynchronous gain fusion mechanism and a dynamic evaluation strategy are incorporated into an UAV autonomous exploration framework, achieving an adaptive balance between “spatial autonomous exploration”and “semantic regularity exploitation”. Simulation results in three Gazebo environments show that the proposed method significantly outperforms baseline approaches in terms of path length, search time, and success rate. In addition, real-world experiments conducted in unknown outdoor environments demonstrate the effectiveness of the proposed method on zero-shot target searching tasks.