The automation of simulation setup—including geometry creation and input file generation—remains a complex and time-intensive task in engineering workflows. In this study, we investigate the potential of Large Language Models (LLMs) to automate this process by translating natural language descriptions into Gmsh-compatible geometry scripts and Elmer simulation input files. We compare the performance of a state-of-the-art off-the-shelf model (LLaMA 3-70B) with a small model (Phi-3 Mini) before and after fine-tuning using Low-Rank Adaptation (LoRA). Using a complex benchmark geometry (wheel-and-axle) and a curated dataset of shapes such as I-beam, we evaluate the models’ capabilities in generating syntactically correct and simulation-ready outputs. Our results show that while large models excel in zero-shot scenarios, fine-tuned small models can match or exceed their accuracy on domain-specific tasks. This approach lowers the barrier for non-expert users and demonstrates the value of lightweight, task-adapted AI in engineering simulation workflows.