DI-engine is a comprehensive reinforcement learning framework that supports distributed training, multi-agent scenarios, model-based RL, and offline RL. It provides implementations of major RL algorithms including IMPALA, R2D2, PPO, and SAC, integrated with PyTorch. The framework includes tools for environment management, exploration-exploitation balancing, imitation learning, and self-play training across diverse domains such as Atari games, MuJoCo robotics, StarCraft Multi-Agent Challenge (SMAC), and MiniGrid.