Ore blending is essential for stabilizing production, reducing costs, and improving resource utilization in mineral processing. This review systematically traces its evolution through four stages: empirical, experimental, mathematical model-based, and systematic integration. Linear programming provides fast, interpretable solutions for clear grade constraints. Nonlinear intelligent algorithms (e.g., particle swarm optimization, genetic algorithms) achieve higher accuracy for multi objective, polymetallic ores but require large datasets and sacrifice interpretability. Systematic blending, powered by 5G, AI, and digital twins, enables full chain coordination with performance gains of>30%, yet requires significant investment. Four interconnected barriers hinder industrial adoption: multi source data integration difficulties, low computational efficiency for real time control, poor model adaptability across ore types (prediction errors up to 10%), and a shortage of interdisciplinary talent. Future directions include ore property based experimental research with open access global databases, hybrid intelligent models with real time dynamic adjustment, and integrated medium to long term blending systems using digital twins and IoT. By explicitly linking experimental data, model calibration, and system deployment, this review provides a structured reference for transitioning from experience driven to data intelligent, full chain synergistic ore blending.