The Institute of Foundation Models, established by MBZUAI, announced PAN 2 on November 10, 2026. The system produces video simulations of the consequences of described actions. Training and inference code and model weights are available. That enables external scrutiny, but does not by itself establish reliable prediction of arbitrary real situations.

Planning and rendering are connected but distinct

The technical account separates a language model that prepares time-ordered states and actions from a video model that constructs the observable scene. The authors report 75.1% completion of individual steps and 34.0% completion of entire plans in their evaluation. These measure different outcomes and are research-team results, not experiments reproduced by SIExpo.

The model card specifies a 29-billion-parameter diffusion component and clips of up to 20 seconds. It identifies prompt sensitivity and consistency loss across clips. IFM's own code and weights use Apache 2.0, while external components retain their respective licenses. Published vLLM recipes use multi-GPU configurations; this is not a promise of easy execution on any home computer.

A convincing final frame is insufficient

Our editorial assessment is that the usefulness of such a tool depends on intermediate states. If an object must move through a doorway, a final image showing it on the other side does not establish that the door opened or that the object passed through it. This is our example of an evaluation criterion, not an additional PAN 2 demonstration.

The distinction between a step and a complete plan has practical consequences. A task may contain several sequential requirements. A good average across individual requirements still leaves room for every complete attempt to contain at least one error. A local-quality number should therefore not be presented as the probability of a successful finished activity. This observation does not require assuming that the errors are independent.

What a useful test would establish

A proposed pilot would begin with actions and observable consequences defined in advance. Teams should preserve the same starting frame, instructions and settings so that outcomes can be compared. Cases in which the intended action should fail are also useful: a simulator that always depicts requested success would be a poor adviser for choosing a plan.

Another check is whether the selected action remains reasonable after receiving a new real observation. Simulation may suggest a hypothesis, but the actual state should take precedence over a persuasive synthetic clip. This is an analytical evaluation framework, not a reported property of a particular deployment.

Evaluation records should also distinguish a generated clip from the decision made using it. A visually better simulation might still lead an agent toward an unhelpful action; conversely, a rough rendering might preserve the particular constraint that matters. The proposed acceptance criterion must specify which outcome the experiment is intended to improve.

Open materials allow more researchers to investigate these questions. The news is an available experimental resource with explicit limitations. SIExpo has not trained PAN 2 or run its GPU evaluations, and does not present it as a certified system for physical control.

Sources: IFM / MBZUAI — PAN 2 release, IFM — PAN 2 technical report and evaluations, IFM — PAN 2 model card and limitations, vLLM — PAN 2 deployment recipe. The cover is a conceptual AI illustration, not a documentary photograph.

Editorial verification: 2026-11-11T17:08:13.509948+00:00.