PAN 2 opens up a world model for simulating actions
IFM released PAN 2's video-model weights and training and inference code for experiments in world modelling.

Before an agent acts, a world model gives it a way to explore what might happen next.
What happened
The Institute of Foundation Models at MBZUAI released PAN 2, which generates video from a starting observation and actions described in text.
The release includes video-model weights and training and inference code under Apache 2.0. Its video backbone has 29 billion parameters and can generate sequences of up to 20 seconds.
Why it matters
Researchers can use accessible code and weights to inspect the system, reproduce experiments and adapt it to new tasks. Simulating candidate actions provides a way to study planning before trying those actions in a physical environment.
The bigger picture
World-model research brings video generation closer to decision-making. Open releases give more teams access to the components needed to investigate how visual predictions can inform an agent's choices.
