To address the high cost of data acquisition, limited sample size, and insufficient prediction accuracy of individual models in blasting fragmentation prediction for open-pit mines, a multi-source fusion dataset comprising 280 samples was constructed through field measurements, literature data collection, and numerical simulations. Blasting design parameters and rock mechanical properties were selected as input variables, and the mean fragment size, X??, was adopted as the prediction target. A blasting fragmentation prediction method integrating multi-source data fusion with Stacking ensemble learning was proposed. Extreme Gradient Boosting (XGBoost), CatBoost, Light Gradient Boosting Machine (LightGBM), and Random Forest (RF) were employed as base learners. The hyperparameters of each base learner were optimized using Optuna, and Ridge Regression was adopted as the meta-learner to achieve weighted fusion of the prediction results generated by the base models. The results showed that a coefficient of determination (R²) of 0.8561, a Root Mean Square Error (RMSE) of 0.0556, and a Relative Absolute Error (RAE) of 0.3241 were achieved by the Stacking ensemble model on the test set, with all evaluation metrics outperforming those of the individual base models. Effective support for blasting parameter optimization and intelligent decision-making in open-pit mines can be provided by the proposed method, while a feasible technical approach to fragmentation prediction under complex geological conditions can also be offered.