<p>The integration of unmanned aerial vehicles (UAVs) as aerial base stations (ABSs) is a pivotal enabler for ubiquitous connectivity in 6G networks. Multiple UAVs have been proposed to serve multiple terrestrial users in a small-cell mobile system. However, optimizing multi-UAV 3D placement to maximize the multi-user system throughput remains a challenging non-convex problem. Existing meta-heuristic algorithms like particle swarm optimization (PSO) offer high precision but suffer from prohibitive computational latency. Conversely, deep learning approaches promise rapid inference but often fail to achieve precise coordinate localization. To bridge this gap, this paper proposes a novel hybrid AI-driven framework combining an adaptive multi-start PSO (AMS-PSO) for high-fidelity data generation and a hybrid U-Net with local refinement for real-time placement. Extensive simulations reveal that the performance of the proposed framework depends on the UAV-to-User density ratio. In scenarios with adequate resources, the hybrid U-Net achieves near-optimal performance, closely matching the AMS-PSO benchmark and significantly outperforming standard heuristics. Even in resource-constrained scenarios, it maintains competitive performance. Furthermore, at high UAV densities, where optimization gains saturate, our method retains a critical advantage in computational speed. Overall, the proposed framework reduces inference time by orders of magnitude compared to iterative heuristics, making it highly viable for dynamic, real-time network orchestration.</p>