Deploy Where You Want: Synthesis of Mobile Network Digital Twins via Generative Artificial Intelligence

High-fidelity wireless mobile network simulation is essential for developing and benchmarking network operational algorithms, yet remains constrained by the high cost of 3D scene acquisition and the lack of joint modeling across the radio, channel, and user aspects. In this paper, we present MobiSim, a deployable mobile network digital twin platform that overcomes these barriers by grounding simulation in generative AI. Given a geographic area, MobiSim automatically constructs a realistic urban environment from public map data and drives a hybrid raytracing engine to model complex multipath propagation, reproducing received-power trends across four real-world measurement datasets spanning sub-6 GHz and millimeter-wave bands. Further, MobiSim co-simulates a complete 5G RAN protocol stack alongside a generative user behavior foundation model that synthesizes realistic mobility trajectories and uplink/downlink traffic demands (app-usage and traffic NMAE reduced by 6.0% and 5.0%, respectively, relative to AppGen on Shanghai), producing high-fidelity full-stack records spanning channel states, infrastructure operations, and user behaviors. The platform further serves as a verifiable reward environment for network optimization, supporting operations research, reinforcement learning, and LLM-based autonomous management. We release MobiSim together with benchmark datasets and standardized evaluation interfaces covering network planning, antenna adjustment, and wireless resource scheduling, providing the community with a reproducible foundation for generalizable mobile network research. The code and generated datasets can be found at https://anonymous.4open.science/r/MobiSim-F533/.