Model–data fusion is defined as matching model prediction and observations by varying model parameters or states using statistical estimation. In this paper, we review the history of applications of various model–data fusion techniques in studies of terrestrial carbon fluxes in two approaches: top-down approaches that use measurements of global CO2 concentration and sometimes other atmospheric constituents to infer carbon fluxes from the land surface, and bottom-up approaches that estimate carbon fluxes using process-based models. We consider applications of model–data fusion in flux estimation, parameter estimation, model error analysis, experimental design and forecasting. Significant progress has been made by systematically studying the discrepancies between the predictions by different models and observations. As a result, some major controversies in global carbon cycle studies have been resolved, robust estimates of continental and global carbon fluxes over the last two decades have been obtained, and major deficiencies in the atmospheric models for tracer transport have been identified. In the bottom-up approaches, various optimization techniques have been used for a range of process-based models. Model–data fusion techniques have been successfully used to improve model predictions, and quantify the information content of carbon flux measurements and identify what other measurements are needed to further constrain model predictions. However, we found that very few studies in both top-down and bottom-up approaches have quantified the errors in the observations, model parameters and model structure systematically and consistently. We therefore suggest that future research will focus on developing an integrated Bayesian framework to study both model and measurement errors systematically.