The Diffusion World Model (DWM) is a conditional diffusion model that predicts multistep future states and rewards simultaneously, enabling long-horizon predictions in a single forward pass. Unlike traditional one-step dynamics models, DWM eliminates the need for recursive queries, reducing compounding errors in long-horizon simulations. DWM is integrated into model-based value estimation, where future trajectories are sampled from DWM to simulate short-term returns. In offline reinforcement learning, DWM acts as a conservative value regularization through generative modeling or as a data source for offline Q-learning with synthetic data. Experiments on the D4RL dataset show that DWM significantly outperforms one-step dynamics models, achieving a 44% performance gain and being comparable to or slightly better than model-free counterparts. DWM is robust to long-horizon simulation, maintaining performance even with a simulation horizon of 31. DWM is also compared with other diffusion-based and transformer-based world models, showing superior performance. The framework includes a diffusion model trained on offline data and an actor-critic method for policy learning. DWM-based algorithms outperform traditional one-step models and are comparable to model-free methods. The approach is flexible and can be extended to various offline learning algorithms. DWM reduces compounding errors by predicting multiple steps at once, leading to better performance in model-based reinforcement learning. The method is effective in offline settings and can be applied to various tasks, demonstrating its potential for practical real-world applications.The Diffusion World Model (DWM) is a conditional diffusion model that predicts multistep future states and rewards simultaneously, enabling long-horizon predictions in a single forward pass. Unlike traditional one-step dynamics models, DWM eliminates the need for recursive queries, reducing compounding errors in long-horizon simulations. DWM is integrated into model-based value estimation, where future trajectories are sampled from DWM to simulate short-term returns. In offline reinforcement learning, DWM acts as a conservative value regularization through generative modeling or as a data source for offline Q-learning with synthetic data. Experiments on the D4RL dataset show that DWM significantly outperforms one-step dynamics models, achieving a 44% performance gain and being comparable to or slightly better than model-free counterparts. DWM is robust to long-horizon simulation, maintaining performance even with a simulation horizon of 31. DWM is also compared with other diffusion-based and transformer-based world models, showing superior performance. The framework includes a diffusion model trained on offline data and an actor-critic method for policy learning. DWM-based algorithms outperform traditional one-step models and are comparable to model-free methods. The approach is flexible and can be extended to various offline learning algorithms. DWM reduces compounding errors by predicting multiple steps at once, leading to better performance in model-based reinforcement learning. The method is effective in offline settings and can be applied to various tasks, demonstrating its potential for practical real-world applications.