The prediction of slope stability is a vital aspect of ensuring safety in geotechnical engineering, as it significantly influences the reliability of infrastructure design and disaster mitigation strategies, particularly in mountainous areas. Traditional prediction methods, such as limit equilibrium approaches and numerical modeling, are often limited by their dependence on simplified assumptions and precise geological parameters, which are notoriously challenging to measure accurately in complex field conditions. Furthermore, these deterministic methods frequently fail to account for the highly nonlinear, heterogeneous, and stochastic nature of slope failure mechanisms. Although machine learning (ML) has emerged as a promising alternative, individual ML models often exhibit limited generalization capabilities, high variance, or bias when applied to diverse geological datasets, resulting in suboptimal predictive performance. To address these ongoing challenges, this study introduces a novel, robust data-driven framework: an Attention-Enhanced Multi-Model Fusion method based on an advanced Stacking ensemble architecture. The proposed system integrates three distinct and complementary base learners—Random Forest (RF), Support Vector Machine (SVM), and XGBoost—selected for their unique capabilities in managing high-dimensional, non-linear geotechnical data. To ensure optimal performance, the hyperparameters of each base learner are meticulously tuned using Bayesian optimization, which effectively navigates the complex parameter space to mitigate the risk of overfitting. The core innovation of this research resides in the development of the meta-learner, which incorporates an advanced dual attention mechanism. In contrast to traditional stacking ensembles that employ static or linear weighting schemes, our attention-based meta-learner dynamically allocates weights to both input features and the outputs of base models. This mechanism enables the model to autonomously identify and focus on critical geotechnical indicators, such as internal friction angle and cohesion, while adaptively prioritizing the most reliable base predictor for specific geological scenarios. Comprehensive experimental evaluations underscore the superior efficacy of the proposed Attention-Stacking framework. The model achieved an impressive accuracy of 0.8604, significantly surpassing standalone base models and traditional stacking methods. A detailed performance analysis reveals a precision of 0.9206, a recall of 0.8169, and an F1-score of 0.8657, indicating an exceptional balance between minimizing false positives and accurately identifying unstable slopes. Furthermore, feature importance analysis highlights the model’s enhanced interpretability; the attention mechanism effectively emphasizes key physical parameters that align with established geotechnical theories, thereby bridging the gap between black-box AI and engineering physics. By concurrently enhancing predictive accuracy, robustness, and transparency, this study presents a reliable, adaptive solution for slope stability assessment, offering substantial potential for advancing risk management protocols and decision-making processes in contemporary geotechnical engineering practices.