Microsoft released MineWorld, the first open-source Minecraft world model capable of real-time interaction at 3+ frames per second. The system uses a VQ-VAE for visual tokenization and a LLaMA-based transformer decoder to predict game states from both video frames and player actions. A novel parallel diagonal decoding algorithm triples generation speed compared to sequential approaches. MineWorld outperforms the closed-source Oasis model on controllability and visual quality metrics while using only 300M parameters, supporting up to 354 actions per minute. The researchers introduced an Inverse Dynamics Model to measure controllability and demonstrated the model's ability to follow action instructions, generate fine-grained details, and maintain contextual consistency across frames.
