<p>Numerical models are essential in the seismic assessment of buildings, and their computational efficiency, predictive capabilities, and reliability are key factors when selecting a modeling approach. However, as simplified representations of reality, models inherently introduce errors and uncertainties in predictions. Despite this, the quantification of uncertainties due to model errors remains largely unexplored, yet it is crucial for informed decision-making and effective risk management in structural assessment. This paper presents the first comprehensive study to quantify model errors and associated uncertainties in the prediction of global seismic response parameters (maximum top floor displacement, maximum base shear, and maximum top floor accelerations) using data from 18 shake table tests on unreinforced masonry buildings subjected to increasing ground motion intensities. The study focuses on the widely used equivalent frame model approach, following consistent and well-defined modeling assumptions. Probabilistic modeling and statistical inference techniques are applied to assess and quantify the model errors and associated uncertainty on the prediction of global response parameters. The key findings are: (1) Model errors in predicting base shear tend to be centered close to zero, bounded between − 35% and + 50%, while errors in displacement and acceleration predictions can be 2 to 3 times larger. (2) Model errors of maximum top floor displacements and accelerations increase with the level of damage, whereas errors for maximum base shear predictions remain relatively constant and are largely unaffected by whether the models are calibrated to match the fundamental period of the buildings. (3) Tuning the Young’s modulus of masonry such that the fundamental period of the model matches the measured fundamental period of the buildings reduces bias and dispersion in predictions for ground motion intensities leading to DG1 or DG2 but it only reduces bias and very slightly the dispersion for ground motion intensities leading to DG3 and DG4. (4) For nearly all buildings, the tuned Young’s modulus is half or less than the Young’s modulus derived from simple compression tests on masonry wallets. Finally, assuming that median errors across different models are equally likely, model uncertainty can be conservatively represented using uniform distributions bounded by the range of median errors for each response parameter and damage level. This provides a practical way to incorporate model error into engineering assessments.</p>