AI & Innovation

PAN 2 world model: MBZUAI releases open AI that predicts physical outcomes

By Anika SharmaAI & Innovations Editor, Peepals Tech
WhatsAppX
Representative image: computer code screen

Representative image: StartupStockPhotos / Pixabay

MBZUAI's Institute of Foundation Models released PAN 2, an open world model that turns a scene and a text action into a predicted video, on 10 November 2026.

What Happened

The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi has released PAN 2, an open "world model" that predicts how a physical scene will change after an action, according to the institute's announcement. Most reports date the release to Tuesday, 10 November 2026. Given a starting scene and an action written in text, PAN 2 produces a video of what it predicts will happen next. Its weights, training code and inference code have been published under the Apache 2.0 licence, which allows free use, including commercial use.

PAN 2: key facts

  • Developer: Institute of Foundation Models at MBZUAI, Abu Dhabi.
  • Released: 10 November 2026, according to most reports (one outlet says 11 November).
  • What it does: takes a scene and a text action and generates a video of the predicted outcome.
  • Size: the released video-generating component is a diffusion model of about 29 billion parameters; a separate language model that plans the next step is not part of the release, according to reports.
  • Licence: Apache 2.0, with weights listed on Hugging Face.
  • Claimed results: third overall and first among open-weight models on the Physics-IQ Verified leaderboard, the institute says.
  • In the institute's own multi-step test, PAN 2 completed 34.0% of full plans against 32.5% for another model, MiniMax-H3, according to reports of its figures.

PAN 2: why it matters

Most well-known AI models work with text or images. A world model tries to do something different: to learn how the physical world behaves, so it can predict what happens when something moves, falls, collides or is pushed. That ability matters for robots, self-driving systems and any AI agent that must act in the real world, because such systems need to test the likely result of an action before taking it.

By releasing the model openly, IFM lets researchers and companies anywhere download it, study it and build on it without paying licence fees. Open releases of this kind let smaller labs, universities and start-ups, including those in India, work with an advanced model without building it from scratch. Diffusion models, the type used here, create images or video by starting from random noise and refining it step by step into a picture that fits the instructions.

There are limits to keep in mind. The benchmark results are the institute's own claims and have not yet been checked independently. One outlet reports that the repository leaves out some training stages used for the public version, while other coverage describes the training code as complete; developers will be able to check the repository for themselves. A score of 34% of full plans also shows how far such systems remain from reliably predicting complex sequences of events.

Peepals calculation: In the institute's multi-step test, PAN 2's 34.0% is 1.5 percentage points above MiniMax-H3's 32.5% (34.0 − 32.5), a relative lead of about 4.6% (1.5 ÷ 32.5). Put the other way, PAN 2 failed to complete about two out of every three full plans in that test (100 − 34 = 66%). Our calculation from the reported figures.

PointPAN 2
DeveloperIFM, MBZUAI (Abu Dhabi)
Released10 November 2026
Released componentAbout 29-billion-parameter diffusion model
LicenceApache 2.0 (open)

Impact

Short-term: Researchers can download PAN 2 and test the institute's claims for themselves, and robotics and simulation teams can try it as a tool for planning actions.

Long-term: Open world models could lower the cost of building AI for robots and physical tasks, and add to competition with closed models from large US and Chinese companies. Their usefulness will depend on how well they predict real-world outcomes outside test conditions.

Who is affected: AI researchers and students, robotics and autonomous-systems developers, start-ups building physical AI, and universities that rely on open models.

Key Takeaway

MBZUAI's Institute of Foundation Models released PAN 2, an open world model that turns a scene and a text action into a predicted video, under the Apache 2.0 licence, with benchmark claims that remain to be independently checked.

Questions and Answers

What is PAN 2?

PAN 2 is an open world model released by the Institute of Foundation Models at MBZUAI in Abu Dhabi, dated 10 November 2026 by most reports. It takes a starting scene and an action written in text and generates a video of what it predicts will happen.

What is a world model in AI?

A world model is an AI system that learns how the physical world behaves, so it can predict the result of actions such as moving or pushing an object. It is useful for robots and other AI agents that must act in the real world.

Is PAN 2 free to use?

Yes. Its weights, training code and inference code are published under the Apache 2.0 licence, which allows free use, including commercial use.

How good is PAN 2?

The institute says it ranks third overall and first among open-weight models on the Physics-IQ Verified leaderboard. These are self-reported results and have not been independently verified.

Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.

#PAN 2#world model#MBZUAI#open-source AI#physical AI

By Anika Sharma, AI & Innovations EditorCovers AI & Innovation and more for Peepals Tech. All stories by Anika Sharma

Get the day's stories in your inbox

Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.

Anika Sharma is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.

Responsible for this story: the Peepals Global Editorial Team, led by editor Prashant Sonar. How we report

More from Peepals Tech

The week's top stories, in your inbox. Free.

The most important, fact-checked stories from all seven publications, once a week.