MBZUAI's institute released PAN 2, an open world model

The Institute of Foundation Models, which was established by Mohamed bin Zayed University of Artificial Intelligence, said on November 10, 2026, that it has released PAN 2, an open world model. In a press release dated November 10, IFM says that if the system is given a starting scene and an action in words, it "simulates a video of what happens next, grounded in physics, time, and space." The institute is releasing video-model weights, full training and inference code, a technical blog and documentation under the Apache 2.0 license, according to that release. This article summarizes the official announcement and same-day coverage. It is not an independent benchmark run or a peer-reviewed evaluation.

Real, reported: based on the public sources cited below. Nothing added or dramatized. Sources checked November 11, 2026, at 7:15 AM ET (EST). Architecture, license and performance claims are the institute's, unless a secondary source is named. Neither the press release nor the coverage we cite describes this announcement as peer-reviewed or names a scientific journal, and we do not do so here.

What IFM says PAN 2 is

IFM describes a world model as a system that simulates the next world, so people and AI agents can ask what would happen if they took an action and see an answer before they act. Researchers at the institute call that "simulative reasoning." Eric Xing, founder of IFM and president of MBZUAI, said: "A language model learns the world as it has been written down. That is book knowledge, and it takes you a long way, but you cannot learn to swim from a book." He added: "Physical intelligence begins when a system can simulate what its actions would do. PAN 2 is our foundation for that."

The press release says PAN 2 combines two kinds of reasoning. In conceptual reasoning, a large language model works out what will happen and when, producing a time-stamped description of events, actions and camera movement. In perceptual reasoning, a video diffusion model, which IFM says it trained from scratch on real-world videos chosen for physical dynamics rather than visual appeal, turns that description into video. The system, IFM says, builds on a Generative Latent Prediction architecture that predicts change in an internal representation and then reconstructs what would be observed.

MarkTechPost, also writing on November 10, 2026, reported that IFM announced the release on X that day and described the video generator as a 29-billion-parameter diffusion transformer. That size figure appears in MarkTechPost's account of the model, not in the press-release text we cite, and should be read as reporting from that article. MarkTechPost also wrote that the system can take text plus an optional first frame and produce 704-by-1248 video at 24 frames per second, for up to 20 seconds.

Reported results, and the institute's own limits

The press release lists three headline results: on six core 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"; on multi-step planning, "PAN 2 completed more full plans than any other model tested"; and on public benchmarks, it was "Third overall on the Physics-IQ Verified leaderboard and first among open-weight models." Those standings are IFM's account of tests it describes. We have not independently verified the leaderboard or reproduced the tasks.

The same official release is explicit about a limit: "Long-term world consistency remains an open challenge, detailed in the technical blog." That caveat belongs beside the headline claims about physics-grounded simulation.

MarkTechPost, labeling the figures as self-reported by IFM, gave additional numbers from the institute's materials, including a 34.0% full-plan completion rate on 100 multi-step tasks, which it said was the highest of the models IFM tested, and it noted that about two in three plans still miss a step. It also reported that IFM's ablation found full-plan completion dropped from 34.0% to 0.0% with no conceptual reasoning. Those details are institute-reported results as relayed by MarkTechPost, not an independent leaderboard we checked. MarkTechPost further wrote that MiniMax-H3 still led on single actions and in a human-preference score in IFM's comparison.

What is being opened, and what is not claimed

Hector Liu, director of IFM's Silicon Valley lab, said: "We want researchers to be able to test PAN 2's capabilities, understand its limits, and build on it. That means releasing the model weights alongside the training and inferencing code." The press release says few world models are released with their full training code, and that PAN 2 is available through IFM's website. Availability of weights and code is not the same as a demonstration that the model is ready for safety-critical robotics, education or real-world control. IFM's own open-challenge statement on long-term consistency cuts against reading the announcement as a finished, general-purpose simulator of the physical world.

MarkTechPost reported that PAN 2 succeeds PAN, which IFM released in November 2025. We do not compare PAN 2's scores with other world-model projects beyond the comparisons IFM itself reported.

Our advice (general guidance)

Read PAN 2 as an institute research release with open weights, not as a certified physics engine or a peer-reviewed ranking. If you download the model or try the demos, use the official IFM pages and the Apache 2.0 license text, and do not treat third-party mirrors or paid "access keys" as official. Simulated videos can look convincing while missing a step, as the coverage of IFM's own plan-completion figures shows, so do not rely on a generated clip as proof of how a machine, a lab procedure or a real environment would behave.

MarkTechPost reported that IFM strongly recommends a separate safety filter and that, in IFM's tests, the filter reduced unsafe outputs on harmful text-to-video requests while also blocking some benign requests. That is the institute's testing as reported by MarkTechPost, not a guarantee. Anyone experimenting with the system should review those recommended safeguards and any applicable local rules on generated media.

These are editorial precautions, not individualized research, engineering or legal advice.

Sources

Published November 11, 2026 by FraudMeNot Team. Label: Real, reported.

Disclaimer. General education only, not legal, financial or security advice. No guarantee. Consult a qualified professional. News summaries are based on the sources listed and may change as more is reported.

General education only, not legal, financial or security advice. Not affiliated with the organizations discussed. No sponsorship or affiliate links. Information may change as further evidence becomes available.