AI-generated concept illustration: possible outcomes of an action in a kitchen evoke world-model simulation.
Before an agent acts, it needs some way to compare what could happen next. A generated video can make a proposed action easier to inspect, but its visual confidence can also conceal a wrong prediction. That tension is central to the practical value of an open world model.
The Institute of Foundation Models, established by the United Arab Emirates' Mohamed bin Zayed University of Artificial Intelligence, released PAN 2 to simulate action outcomes as video. IFM's own release is dated November 10 in Sunnyvale, California; the university posted its announcement on November 11. This is a release from earlier this week, not a new November 13 launch. [P1][P2]
The public repository describes optional image-and-text observations and text-described actions producing videos of up to 20 seconds. Its architecture combines a language model that describes events over time with a roughly 29-billion-parameter diffusion backbone that renders the predicted changes. The aim is to let agents compare candidate actions through simulation. [P3]
For a researcher, that creates a concrete test: does selecting an action with the simulator produce a better real outcome than choosing without it? The test needs measured consequences and a baseline. A compelling clip of a kitchen task cannot establish that an unfamiliar robot, surface or object will behave the same way.
The release offers PAN 2 code and weights under Apache 2.0, with separate terms for third-party components. There is an important qualification to the broad openness language in the announcements: the repository says its training code covers pretraining and supervised fine-tuning, while reward learning and the model-soup procedure used to produce the released weights are described in the technical report and are not included in that code. [P1][P3]
Those are different levels of reproducibility. Using released weights to test a prompt is one exercise; rebuilding the final released model from a complete training pipeline is another. Readers evaluating the release should keep the distinction visible rather than treat an open repository as proof that every stage can be rerun.
MBZUAI acknowledges that maintaining world consistency over long periods remains an open challenge. Its performance comparisons are developer-reported results; AIYXL has not run the model or independently reproduced those evaluations. The confirmed development is the public research release and its documented scope, not a demonstration of production-ready autonomous control. [P2][P3]
PAN 2's significance is therefore an expanded opportunity to investigate action simulation. Useful follow-up work would record where simulated outcomes diverge from observations, whether errors compound across steps and when a planner should decline to trust a prediction. An open model can make those questions easier to study without resolving them in advance.
Sources:
