Discrepancies between future projections of land carbon fluxes originate from different process representations, but also from differences in model parameterization. Model parameters are typically drawn from disparate literature sources, individual site measurements, or expert judgment, allowing for large variability in the actual parameters used and thus model functional responses. In addition, differences in meteorological forcing datasets and modeling setups may contribute strongly to differences at inter-annual and longer time scales. Along with the known differences in the mean model behavior under future climate scenarios, there are likely also differences in model responses to increased climate variability, and extreme events, which have yet to be assessed. In this study we developed an in situ model data fusion experiment to explore the contribution of diverse long-term observations in addressing the divergence of modeled projections of ecosystem water and carbon fluxes until 2100, along with responses to climate variability and extreme events. We focus on two forest sites in France – Hesse and Le Bray – for which carbon and water fluxes have been observed for more than ten years using eddy covariance methodology. The consolidated set of eddy covariance observations and respective uncertainties is complemented with biometric information on aboveground biomass, biomass increments and soil carbon stocks. These datasets are simultaneously used as constraints in the inverse parameter optimization of an ensemble of terrestrial biogeochemical models ranging from specific forest models to generic land surface schemes, namely: BASFOR, FöBAAR, JSBACH, LPJ and ORCHIDEE. The experimental setup includes the harmonization of the optimization by forcing and constraining the models with the same observations, and through a common cost function. The set of multiple constraints ensures that the models simulate the responses of ecosystem fluxes to environmental conditions in agreement with ecosystem pools. In all models we observe significant improvements in modeling performance but modest improvements in estimating the interannual variability in carbon fluxes and pools. The divergence in long-term trends until 2100 between models is reduced in the carbon fluxes and pools after optimization. However, an increase in the variability of net ecosystem fluxes is observed, which results from the higher interannual variability in the climate scenarios, as well as the growing ecosystem carbon pools. These results suggest more frequent and amplified responses of ecosystem carbon cycle as present-day extreme conditions become more frequent. Overall, this study emphasizes the importance of long-term observations in assessing inter-model divergence and in addressing the future sensitivities of ecosystem carbon fluxes to changes in climate variability.