MBZUAI Releases PAN 2, an Open-Source World Simulation Model

How it works

PAN 2 combines two levels of reasoning. In conceptual reasoning, a large language model determines the sequence of events and actions, camera movement and expected timing. In perceptual reasoning, a diffusion model built for video turns those descriptions into scenes with motion details and physical interactions.

The institute says it trained the video model from scratch on real-world clips chosen for their physical dynamics rather than visual appeal. PAN 2 also uses a Generative Latent Prediction (GLP) architecture, which predicts changes in an internal representation of the world before reconstructing the visual output.

Results and limits

The institute says PAN 2 ranks third overall on the Physics-IQ Verified leaderboard and first among open-weight models. It also reports strong results across six dimensions of physical intelligence, including causal consistency, entity persistence and object dynamics, and says it outperformed the other models in the comparison on a multi-step plan completion test.

The model can simulate scenarios absent from its training data, such as an egg falling onto different surfaces, and the virtual environments it produces could be used to train agentic AI systems to explore and plan. The institute acknowledges that keeping a simulated world consistent across long, multi-step interactions remains an open research challenge.

PAN 2 is released under the Apache 2.0 license, with the video model weights, training and inference code, and a technical blog post. Hector Liu, director of the institute's Silicon Valley research lab, said releasing these components is meant to let researchers test the model's capabilities and limits and build on it. It is available through the institute's website and can be adapted for research in areas such as robotics and education.