The rapid development of big data has shifted the traditional decision-making to data-driven decision-making. This research aims to propose an O2O on-demand logistics management model based on the new big data-driven paradigm. This model incorporates different data sources from the internal logistics departments, other operations departments and external environment to form a panoramic dataset. Based on this, the premised assumptions of the traditional decision-making paradigm can be relaxed, and two assumption transformations are realized: from the consistent delivery time to personalized delivery time, and from the prior demand distribution to the temporal-spatial demand distribution. This research aims to realize the personalized O2O on-demand delivery management by applying both machine learning and operations research technologies. More specifically, a personalized delivery time forecast model and a scenario-based demand forecast algorithm are proposed. The point estimate and forecast uncertainty of the delivery time forecast model, and the future temporal-spatial demand distribution of the demand forecast model are incorporated into the decision-making model for O2O on-demand logistics systems. A forecast-while-optimize algorithm is also developed to optimize the decision-making model based on the feature-dependent predictions. This research verifies the feasibility and effectiveness of the O2O on-demand logistics management model based on the new paradigm by analyzing a real dataset from one of the largest O2O platforms in China. Compared with the traditional mode, the model based on the big data-driven paradigm can precisely match the highly uncertain demand and supply, and reduce the number of delayed orders, average delivery time and delivery cost.