Humanoid robots are highly valued for their human-like operation and adaptability to human environments but face limitations in terms of passability in complex outdoor terrains and moving speed. Conversely, aerial robots offer superior high-speed mobility and obstacle navigation. Flying humanoid robots, merging human-like adaptability with aerial mobility, have emerged as a highly promising frontier in robotics, offering innovative solutions for efficient navigation and manipulation tasks in complex environments. However, existing flight control methods often rely on single-rigid-body models, lacking the capability to dynamically update varying flight parameters such as center of mass (CoM) position and inertia. Conventional approaches also struggle with thrust magnitude-angle coordination and actuator constraints, risking output saturation. This paper presents a Thrust-Vector Model Predictive Control (TV-MPC) method for a flying wheel-legged humanoid robot (FWLR), addressing key challenges in flight control. We establish a Vector Thrust Projection Model (VTP) based on centroidal dynamics, enabling a linearized description of the robot’s flight dynamics. Leveraging MPC’s receding horizon optimization, our approach achieves adaptive model parameter updates, decoupled position-attitude control, optimized trajectory tracking, and optimal thrust intensity-angle allocation. Furthermore, actuator constraints are seamlessly integrated into the MPC framework through linearized thrust-vector formulations. Simulation experiments demonstrate that the proposed method achieves decoupled position-attitude flight control for FWLR, enabling precise tracking of 6-DoF trajectories while constraining actuator outputs within feasible limits.