This paper presents GUMP, a scalable generative model for motion planning in autonomous driving. The model learns the dynamics of driving scenes and can simulate diverse future scenarios and generate various driving scenarios based on prompts. It operates in both full-autoregressive and partial-autoregressive modes, improving inference and training speed without sacrificing generative capability. The model is evaluated on the Waymo motion dataset and the nuPlan dataset, achieving state-of-the-art performance in simulation realism and scene generation. It outperforms prior arts in planning benchmarks. The model serves as a foundation for various motion planning tasks, including data generation, simulation, planning, and online training. GUMP is designed with a key-value pair tokenizer and a Multimodal Causal Transformer (MCT) to predict future scenarios. It enables efficient scene generation, scene extrapolation, planning, and reinforcement learning. The model demonstrates high scalability and strong generalization capabilities, and is used as an online training environment for reinforcement learning. The model's performance is validated on multiple benchmarks, showing significant improvements in simulation realism and planning accuracy. The model is also effective in interactive planning tasks and online training, demonstrating its potential as a foundation for autonomous driving systems.This paper presents GUMP, a scalable generative model for motion planning in autonomous driving. The model learns the dynamics of driving scenes and can simulate diverse future scenarios and generate various driving scenarios based on prompts. It operates in both full-autoregressive and partial-autoregressive modes, improving inference and training speed without sacrificing generative capability. The model is evaluated on the Waymo motion dataset and the nuPlan dataset, achieving state-of-the-art performance in simulation realism and scene generation. It outperforms prior arts in planning benchmarks. The model serves as a foundation for various motion planning tasks, including data generation, simulation, planning, and online training. GUMP is designed with a key-value pair tokenizer and a Multimodal Causal Transformer (MCT) to predict future scenarios. It enables efficient scene generation, scene extrapolation, planning, and reinforcement learning. The model demonstrates high scalability and strong generalization capabilities, and is used as an online training environment for reinforcement learning. The model's performance is validated on multiple benchmarks, showing significant improvements in simulation realism and planning accuracy. The model is also effective in interactive planning tasks and online training, demonstrating its potential as a foundation for autonomous driving systems.