Dexterous in-hand manipulation is an essential skill of production and life. Nevertheless, the highly stiff and mutable features of contacts cause limitations to real-time contact discovery and inference, which degrades the performance of model-based methods. Inspired by recent advancements in contact-rich locomotion and manipulation, this paper proposes a novel model-based approach to control dexterous in-hand manipulation and overcome the current limitations. The proposed approach has the attractive feature, which allows the robot to robustly execute long-horizon in-hand manipulation without pre-defined contact sequences or separated planning procedures. Specifically, we design a contact-implicit model predictive controller at high-level to generate real-time contact plans, which are executed by the low-level tracking controller. Compared with other model-based methods, such a long-horizon feature enables replanning and robust execution of contact-rich motions to achieve large-displacement in-hand tasks more efficiently; Compared with existing learning-based methods, the proposed approach achieves the dexterity and also generalizes to different objects without any pre-training. Detailed simulations and ablation studies demonstrate the efficiency and effectiveness of our method. It runs at 20Hz on the 23-degree-of-freedom long-horizon in-hand object rotation task.