Abstract. Ecosystem CO₂ fluxes in high Arctic are rather dynamic, as they are sensitive to climatic variability through multiple ecosystem processes, for instance, vegetation and snow dynamics as well as permafrost thawing, operating at different time scales. Uncertainties from both high-frequency measurements and model assumptions challenge model calibration to describe both short- and long-term phenomena related to weather and climate variabilities. In this study, we generated three model ensembles using a Monte-Carlo based uncertainty approach with acceptance criteria for 15 years of eddy covariance CO₂ measurements of a high Arctic heath ecosystem based on the time-integrated CO₂ fluxes within the day, the year and the entire period. The temporal distribution of residuals between the model and measurements indicated that the three model ensembles reasonably simulated diurnal, seasonal and long-term behaviours of CO₂ fluxes respectively. The inter-annual variation of CO₂ fluxes over 15 years showed the current ecosystem is at a transition from being a C sink to a C neutral balance. The long-term behaviour model ensemble simulated a more intensified diurnal C cycle than the short-term behaviour model ensembles. The intensified C cycle was mainly attributed to a faster depletion of the soil C pools. The sensitivities of posterior parameters to the model performance index (coefficient of determination, R²) reflected that parameters in the processes of soil water and heat transfer and snow dynamics regulated the short-term behaviour of CO₂ fluxes, while parameters in the process of soil decomposition regulated the long-term behaviour of CO₂ fluxes. Our results suggest that the development of ecosystem models should diagnose their effectiveness in capturing ecosystem CO₂ exchange behaviour across different time scales. A clear trade-off may exist when the model is tuned to capture both the short- and long-term variation of CO₂ fluxes. To constrain the model with the time-integrated CO₂ fluxes is a simple and useful method to reduce the non-explained errors and to identify the crucial link to controlling parameters and processes.