Abu Dhabi’s MBZUAI releases PAN 2, an open physics-grounded world model

The latest:

A world model that simulates the physical consequences of an action is now publicly available, after the Institute of Foundation Models (IFM) released PAN 2 with its weights and full training code. Founded by Mohamed bin Zayed University of Artificial Intelligence, IFM said the model takes a starting scene and a described action, then generates video of the outcome grounded in physics, time and space.

Details:

• The concept: IFM frames PAN 2 around what its researchers call simulative reasoning: where a language model predicts the next word, a world model simulates the next world. The goal is to let people and AI agents ask what happens if they act, and see the answer before committing to costly or irreversible real-world trials.
• The architecture: According to IFM, the system pairs two forms of reasoning. A large language model handles conceptual reasoning, producing a time-stamped description of events, actions and camera movement. A video diffusion model then converts that description into video, filling in details such as how a wave breaks or how a hand grips a pipette.
• The training data: IFM said it trained the video diffusion component from scratch on real-world videos selected for their physical dynamics rather than visual appeal, a choice aimed at physical fidelity over aesthetics. The institute did not disclose the size of that dataset.
• The underlying method: PAN 2 builds on the Generative Latent Prediction architecture developed by MBZUAI President Eric Xing and IFM researchers. Per IFM, the approach predicts change inside an internal representation and then reconstructs what would be observed, so predictions can be checked against the real world rather than taken on trust.
• The benchmarks: On six core dimensions including physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamics, IFM said PAN 2 placed just behind MiniMax-H3 and ahead of four other models. On multi-step planning, it reported PAN 2 completed more full plans than any other model tested.
• Public leaderboard: IFM said PAN 2 ranked third overall on the Physics-IQ Verified leaderboard and first among open-weight models. The institute did not name the models placing above it there.
• Beyond the data: The institute said PAN 2 can simulate situations absent from its training data, citing an egg dropped onto a wide range of surfaces, and can act as an environment where AI agents explore and learn, such as cooking pasta in a kitchen they have never seen before. IFM acknowledged that long-term world consistency remains an open challenge.
• The release terms: IFM is publishing video model weights, full training and inference code, a technical blog and documentation under the Apache 2.0 license, noting that few world models are released with their full training code. Hector Liu, director of IFM’s Silicon Valley lab, said the aim is for researchers to “test PAN 2’s capabilities, understand its limits, and build on it.”
• The pitch: Xing argued that a language model learns the world as it has been written down, which he described as book knowledge that takes you far but cannot teach you to swim. He said physical intelligence begins when a system can simulate what its own actions would do.

Between the lines:

The competitive claim here is less about the leaderboard placements than the licensing. IFM itself notes that few world models ship with full training code, so releasing weights, training and inference code under Apache 2.0 positions PAN 2 as the reproducible option rather than the top-scoring one — consistent with a model that ranked third overall but first among open-weight entrants.

What’s next

Watch whether independent researchers reproduce the Physics-IQ Verified and multi-step planning results now that the training code is public, and whether robotics and education teams publish adaptations built on the Apache 2.0 release.