A Guth Labs publication ↗
Back to News

IFM releases PAN 2 world model with training code and open weights

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence announced PAN 2 on November 10, releasing a world-model system intended to simulate the consequences of actions before an agent takes them. IFM says its own code and video-model weights are available under Apache 2.0. IFM announcement

PAN 2 combines a language model that describes what should happen and when with a 29-billion-parameter diffusion model that generates the visual result. Given text and an optional starting image, it can simulate a video lasting up to 20 seconds. The public repository contains training and inference code, while separately supplied components retain their own licenses. PAN 2 repository

The developers report 75.1% completion of individual steps and 34.0% completion of full plans in their evaluation. Those are materially different measures: success on many isolated actions can still leave a long sequence unfinished. The comparison also uses different language-model backbones across benchmark settings, so its results should not be reduced to a single universal ranking. Technical report

IFM’s announcement acknowledges that maintaining a consistent simulated world over long periods remains an open problem. The project website lists a research paper as forthcoming; the current technical blog and released implementation provide the available detail. IFM announcement Project page

The release gives researchers a concrete system to inspect and adapt for experiments in action planning. Its value for physical deployment will depend on whether simulated outcomes match real observations, including unusual conditions and failures. Plausible video is useful evidence only when the predicted consequences survive that comparison.

Sources and citations

Revision history

Published .