PAN 2: An Open World Model That Simulates What Happens Next

PAN 2: An Open World Model That Simulates What Happens Next
AI-generated illustration

A language model predicts the next word. A world model tries to predict the next moment of the world itself. On November 10, 2026, the Institute of Foundation Models (IFM), established by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), released PAN 2, an open, general-purpose world model aimed at physical intelligence.

The pitch is simple to state: give PAN 2 a starting scene and an action described in words, and it simulates a video of what happens next, grounded in physics, time and space.

Simulative reasoning, in plain terms

IFM researchers describe the idea as simulative reasoning. Instead of trying something in the real world, which can be slow, costly or irreversible, a person or AI agent asks “what happens if I do this?” and watches the answer first. Eric Xing, IFM’s founder and MBZUAI’s president, put it this way: a language model learns the world as it has been written down, but you cannot learn to swim from a book.

Two kinds of reasoning in one system

According to IFM’s release, PAN 2 pairs two components. The first is conceptual reasoning: a large language model works out what will happen and when, producing a time-stamped description of events, actions and camera movement. The second is perceptual reasoning: a video diffusion model, trained from scratch by IFM on real-world videos chosen for physical dynamics rather than visual appeal, turns that description into video. It fills in details such as how a wave breaks or how a hand grips a pipette.

PAN 2 builds on the Generative Latent Prediction (GLP) architecture pioneered by Xing and IFM researchers. GLP predicts change in an internal representation, then reconstructs what would be observed, so predictions can be checked against the real world.

The results IFM reports

IFM highlights three results. On six core dimensions, covering 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. On multi-step planning, it completed more full plans than any other model tested. And on the public Physics-IQ Verified leaderboard it ranked third overall and first among open-weight models.

The model can also simulate situations missing from its training data, such as an egg dropped onto a wide range of surfaces. It can act as a world in which AI agents explore and learn, for example cooking a pot of pasta in a kitchen they have never seen. IFM is candid about one limit: long-term world consistency remains an open challenge, detailed in its technical blog.

Open weights and open code

The release includes video model weights, full training and inference code, a technical blog and documentation, all under the Apache 2.0 license. IFM notes that few world models ship with their full training code. Hector Liu, director of IFM’s Silicon Valley lab, said the goal is for researchers to test PAN 2’s capabilities, understand its limits and build on it. IFM says the model is available now through its website and is meant to be adapted for fields from robotics to education.

IFM itself launched in May 2025 under MBZUAI, with labs in Abu Dhabi, Silicon Valley and Paris. Its portfolio includes the K2 series of open-source language models, the Arabic model Jais and the PAN world-model line.

What it means

For robotics and agent researchers, a world model with open training code is a sandbox for testing plans before touching hardware. For educators, simulated scenarios could make physical processes easier to demonstrate. For the wider open-source community, it is a rare chance to inspect and reproduce a world model end to end.

Keep expectations measured. The rankings are IFM’s reported results, a simulated video is not a guarantee of how the real world behaves, and long-horizon consistency is still unsolved.

Conclusion

PAN 2 pushes the world-model idea from closed demos toward something anyone can download, probe and improve. How well its simulations transfer to real robots and real classrooms is the next question to watch.

Sources