The Institute of Foundation Models released PAN 2 on November 10, describing a world model that turns a starting scene and a written action into a video of what might happen next. The institute's announcement says it is releasing the weights, training code, inference code and documentation so researchers can inspect and extend the model.
How does PAN 2 make a prediction?
The model uses an architecture called Generative Latent Prediction. It first predicts how an internal representation of the scene will change. It then reconstructs the predicted observation as video. That lets a researcher compare what the model produced with what can be observed in the real world, instead of judging only whether the clip looks plausible.
The institute illustrates the idea with examples such as dropping an egg on different surfaces or simulating an activity in a kitchen the agent has not seen. These are examples the developer uses to describe the model's purpose. They are not a guarantee that a generated video gets every physical interaction right.
What can researchers download?
The institute says the release includes the model weights, full code for training and inference, a technical blog and documentation under the Apache 2.0 license. The training code matters for readers who want to examine how the model was built, rather than evaluate it only through a demo. The release describes a model for simulating sequences after an action, not evidence that it has been validated for a particular deployment.
Where does the simulation still fall short?
The institute calls long-term world consistency an open challenge. A plausible change at one step does not establish that objects and conditions will remain consistent across a longer sequence. A useful first evaluation is to compare a short prediction with a known real outcome, then check what stops matching as the sequence gets longer.
