This article presents a comprehensive investigation on the scalability of a state-of-the-art air pollution model, analysing its performance across varying high performance computational resources. Scalability refers to the model’s ability to efficiently utilize certain amount of scalable computational resources as the size of the simulation domain and/or the complexity of the model increase too. Parallel algorithms are among the key techniques in achieving high scalability. Efficient data partitioning is another key issue. Sensitivity studies in air pollution modeling are essential for understanding the reliability and robustness of these models and for informing decision-making processes related to air quality management and environmental policy. These are aimed at investigating how sensitive a complex air pollution model is to variations or uncertainties in its input data, parameters or assumptions. The Unified Danish Eulerian Model (UNI-DEM) is the particular air pollution model, discussed here. It calculates the concentrations of a large number of pollutants and other chemical species in the air and their variation along certain time period, taking into account the main physical, chemical and photochemical processes in the atmosphere. Scalability properties of the computer implementation of the model and its sensitivity analysis code (SA-DEM), run on the most powerful Bulgarian supercomputer—Discoverer, are studied. One possible scheme for metamodelling of a such kind of a large-scale mathematical model have been described. Several sensitivity analysis techniques and their advantages and disadvantages are analysed too. Several efficient Monte Carlo algorithms for computing small in value sensitivity measures (that are crucial for preparing the final predictions on the base of sensitivity analysis results) have been applied and compared. At the end, some results of sensitivity study of UNI-DEM output mean monthly concentrations with respect to the input emission levels are presented.