Abstract: Frequent occurrences of slope-toe landslides in southwestern mountainous areas of China often endanger life and property in slope-front residential zones. The energy dissipation at turning points of slope-toe landslides substantially increases the difficulty in predicting its maximum horizontal movement distance (L_max). To address this challenge, this study develops a BP neural network-support vector machine ensemble model integrated with quantile regression framework (QR-BP-SVR), incorporating six parameters: landslide volume (V), slope gradient (α), et al. The model successfully predicts L_max for slope-toe landslides under small-sample conditions. Comparative analyses with BPNN, SVR, Random Forest (RF), and Multilayer Perceptron (MLP) demonstrate that the ensemble model achieves significant improvements in key metrics, including R² (mean 0.9146) and MAE (mean 57.71 m). SHAP value analysis and partial dependence plots identify V, L₁, H₁ as primary controlling factors of L_max, while revealing inflection point effects of f and γ. The quantile regression-enhanced ensemble model effectively reduces prediction errors, offering a robust solution for L_max prediction of slope-toe landslides in complex terrains under limited data conditions.