Could time series foundation models improve the forecasting of district heating demand? In our offline study, we compared Chronos-2, TimesFM-2.5, and Moirai-2 with AutoARIMA, XGBoost, and other baselines across four real district heating grids. Chronos-2 produced the most accurate results, even in zero-shot settings without target-grid training data. We also examine data efficiency, fine-tuning, resilience to faulty sensor inputs, and operational reliability.