Alcohol use disorder (AUD) is among the most prevalent mental disorders. Associations between resting-state fMRI functional connectivity (FC) and AUD have been reported but rarely replicated. In this preregistered study, we investigated the extent to which FC predicts alcohol consumption, as measured by the AUDIT-C score, in a large sample of non-treatment-seeking alcohol consumers (N = 739), including 528 participants with varying degrees of AUD severity. To optimize performance, we compared kernel ridge regression (KRR) trained directly on this sample with transfer-learning models based on multilayer meta-matching embeddings, from which we derived data-driven relevance maps at the connection, region, and network levels. We tested split-half reliability, reproducibility across cohorts, and outcome specificity of the relevance maps. Both models predicted AUDIT-C scores moderately but significantly (r ≈ 0.15). Meta-matching did not outperform KRR in prediction. Relevance maps showed limited split-half reliability, consistent with the modest predictive performance, but reproduced moderately well across two independent cohorts. These maps indicate reduced connectivity among limbic regions, the anterior middle temporal gyrus, the dorsomedial prefrontal cortex, and dorsal attention network regions. In contrast, hypothesized hyperconnectivity patterns involving the striatum, anterior cingulate cortex, and precuneus were less reliable and mostly non-significant. Overall, our findings suggest that resting-state FC contains reproducible but weak signals related to alcohol consumption. Transfer learning improved relevance-map reliability but not prediction performance and reduced outcome specificity. These results support the utility of interpretable machine learning for identifying connectivity patterns related to drinking behavior, while also highlighting current limitations in its predictive power.