Decadal climate prediction offers critical scientific guidance for policymakers and supports socioeconomic sustainability. Nevertheless, state-of-the-art dynamical models remain limited in predicting the intricate multidecadal variability in Eurasian surface air temperature(SAT) at mid-to-high latitudes several years in advance. Here, we present a hybrid deep learning model–Gated Recurrent Unit augmented with batch normalization and attention mechanism (GRUBA)–that significantly enhances decadal prediction skill through advanced postprocessing of multi-model ensemble outputs. By incorporating K-means temporal clustering, GRUBA sequentially refines each SAT cluster accordingto its distinct decadal variability. During the test period of 2004–2021, the averaged anomaly correlation coefficient skill improves from –0.23 to 0.83, and the mean square skill score from –0.37 to 0.68 compared to the unweighted multi-model ensemble mean. This substantial improvement arises in part from the attention mechanism, which effectively reduces model spread. Furthermore, SHAP (Shapley Additive exPlanations) analysis reveals that GRUBA assigns higher weights to better performing ensemble members.