Solving engineering problems requires a solid understanding of the physical world. In particular, tasks involving MultiBody Systems (MBSs) with flexible bodies demand the capability to handle three-dimensional geometries, apply Finite Element Methods (FEMs), and perform spatial reasoning. Although generative language models excel across disciplines, they remain limited in applying FEMs and geometry handling. The objective and novelty of this work lie in the automated generation of flexible MBSs with systematic validation without finetuning Large Language Models (LLMs). In the recent work, a subset of the authors proposed a pipeline for LLM-based, automated generation and validation of simulation code for simple MBSs with mass points and rigid bodies using the open source code Exudyn. In line with this latter work, we adopt an approach that extends the model generation pipeline and streamlines the Exudyn API by reducing options and simplifying functionalities. Our approach decomposes MBS model generation into three stages: determination of body geometries and required flexible bodies; generation geometry and mesh for flexible bodies, modal reduction, and interface definitions using Netgen/NGSolve; and automated assembly of complete simulation models integrating flexible and rigid bodies, joints, loads, and connectors. For each task, the LLM identifies the required simulation elements, after which the pipeline constructs specialized context from documentation and code snippets that the LLM uses to generate the simulation code. Outputs are benchmarked against reference implementations under harmonized solver settings. Executability and physical correctness are distinguished: the former requires full model execution, the latter bounded deviation from ground truth responses, including inertia properties and eigenfrequencies of flexible bodies. Preliminary results demonstrate successful automated construction of flexible multibody systems beyond academic examples, including interacting flexible bodies in rotor-support configurations. The pipeline is training-free, transferable to other multibody codes, and highlights the near-term potential of LLMs to model and simulate real machines as flexible MBSs.