Buildings account for around 40% of final energy use in the EU, and most savings from retrofits, better control and demand flexibility depend on calibrated building energy models. Building such models in Modelica, EnergyPlus or IDA ICE is slow, expert work, so most existing buildings are never modelled at all. This project investigates whether open-weight large language models (LLMs), run locally as autonomous agents, can produce building energy models that are physically valid, calibrated to measurements and useful for energy optimisation. The agent works from heterogeneous building information and iteratively builds and revises simulation models using open-source modelling tools. We study how different agent designs and levels of engineering guidance affect the quality of the resulting models, using public benchmark buildings on NAISS resources and, on local hardware only, real buildings with measured data. Models are assessed for accuracy against measurements and reference models, physical plausibility, and reproducibility across repeated runs. Open-weight models are essential here: operator data from real buildings cannot be sent to commercial APIs, and fixed model weights make the experiments reproducible. We evaluate models from about 8B to about 670B parameters to quantify how model size affects modelling quality, and compare them with small domain-adapted models (1–8B) fine-tuned on local hardware. A price-signal heating optimisation case quantifies the energy and cost savings the generated models enable. Principal Investigator: Zeng Peng, doctoral student, KTH Royal Institute of Technology. Supervisor: Qian Wang, KTH Royal Institute of Technology, Department of Civil and Architectural Engineering. Thomas Ohlson Timoudas, RISE Research Institutes of Sweden, Division of Digital Systems, Computer Science.