Water and carbon exchanges are coupled because carbon dioxide (CO 2 ) and water vapor loss occur through leaf stomata, yet land surface predictions remain uncertain because vegetation and soil parameters are poorly constrained and often calibrated independently at individual sites. We develop a hierarchical ensemble smoother with multiple data assimilation (H-ES-MDA) for the enCore Community Land Model (eCLM) that pools information across 38 European eddy-covariance sites. The framework estimates shared plant functional type (PFT) vegetation parameters while retaining site-level vegetation deviations and site-specific soil hydraulic parameters. Six assimilation scenarios combined soil moisture (SM), evapotranspiration (ET), net ecosystem exchange (NEE), with performance assessed in independent validation years. Joint assimilation of SM, ET, and NEE produced the most balanced improvements relative to prior simulations, improving median validation performance by 30–46% for SM, 7% for ET, and 18% for NEE, while avoiding the SM degradation observed with flux-only assimilation. Shared PFT posteriors retained substantial NEE skill for grasslands and evergreen needleleaf forests, whereas site-level deviations added little skill. A leave-one-site-out experiment for grasslands confirmed that shared NEE skill persisted at sites excluded from calibration. Transfer was weaker for deciduous broadleaf forest and cropland and did not consistently extend from NEE to unassimilated gross primary production and ecosystem respiration. Kullback–Leibler (KL) divergence and Shapley attribution further showed that strong parameter updating does not necessarily imply predictive skill. Hierarchical multi-site assimilation therefore provides a direct way to test which parameter information can be shared regionally and which processes require local or improved model representation.