MBZUAI releases PAN 2

The Institute of Foundation Models at MBZUAI released PAN 2 on 11 November 2026. The press release calls it an open world model. You give it a starting scene and an action in words, and it simulates a video of what happens next. The release describes two steps. A language model writes a time-stamped account of events, actions, and camera movement. A video diffusion model, trained from scratch on real-world video chosen for physical dynamics, turns that account into pictures.

The release says that on six dimensions, physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance, and objective dynamics, PAN 2 placed just behind MiniMax-H3 and ahead of four other models. It says PAN 2 completed more full plans than any other model in its multi-step planning test. On the public Physics-IQ Verified leaderboard, the release says PAN 2 is third overall and first among open-weight models. Those placements are the institute's. It also says long-term world consistency is still an open problem, and points to a technical blog for that limit.

The release includes video model weights, full training and inference code, and documentation, under the Apache 2.0 license. The institute says the model is available through its website.

Who it's for: a researcher who wants an open world model, with training code, to test a simulated consequence before acting.

Fits: Best for Video stays Runway, for finishing a shot with camera moves and inpainting. PAN 2 simulates a consequence. It leaves that editing job where it is. The video guide is unchanged. The verdict is unchanged.