What happens after an object is pushed across a table? Answering that question involves position, motion, contact and timing. For an AI system planning a sequence of actions, small mistakes in those details can change the entire outcome.

PAN 2 explores this problem through video simulation. Released on November 10, 2026, by the Institute of Foundation Models (IFM), established by MBZUAI, it gives researchers a way to examine how a proposed action might unfold.

The public release is useful because it gives developers something concrete to investigate: downloadable weights, code and documented limitations. Whether the simulations help a particular application remains a question for testing.

How a scene becomes a simulation

PAN 2 takes a description of a scene and actions, with an optional starting image, and produces a video of up to 20 seconds.

A language model first organizes the request into a structured account of events, actors, timing and camera movement. A text encoder processes that description. The video diffusion backbone then predicts a latent representation of the scene’s development, and a video decoder turns it into visible frames.

For someone evaluating the system, this creates several places to look when an output goes wrong. The action may have been described ambiguously. The structured prompt may have lost an important condition. The video stage may have failed to preserve an object’s position or identity. Keeping the intermediate description makes those possibilities easier to investigate.

What is available to download?

The released diffusion backbone has 29 billion parameters. One checkpoint supports both text-to-video and image-to-video generation. IFM provides the weights through Hugging Face and training and inference code through its GitHub repository.

A complete setup also uses separately distributed components, including the Qwen3.5-9B text encoder and Wan 2.2 VAE. The language model responsible for structuring the scene and action instructions must also be accounted for.

PAN 2’s own code and weights carry the Apache 2.0 license. Third-party dependencies retain their respective licenses. Anyone packaging a service should check the notices and licensing requirements for the full set of components they intend to use.

What to check before deployment

The repository documents environment setup, checkpoint conversion and inference. With its standard inference launcher, each GPU worker holds the full 29B model and processes a share of the prompts. Extra GPUs increase the number of requests that can be processed alongside one another.

The inference documentation also describes offloading options that move models or blocks between CPU memory and the GPU. Their practical value depends on the hardware and configuration. We have not deployed PAN 2 for this article, so we cannot provide a tested minimum GPU specification or generation-time estimate.

A sensible first trial would use a small, fixed set of inputs. Save the starting image, rewritten prompt, configuration and output together. That makes it easier to compare changes and to reproduce a failure after adjusting the settings.

Where the limitations matter

IFM’s model card acknowledges substantial problems with longer plans, prompt sensitivity and consistency across clips. It also describes cases where the simulation skips a necessary physical step while still moving toward the requested outcome.

That last point deserves attention. A convincing final frame can conceal a faulty sequence. When reviewing an action, follow the objects through the whole clip. Check their locations, the order of events and the conditions that must hold before each step can happen.

Useful evaluation questions include: Does the same object remain visible throughout the action? Does its final position follow from its earlier movement? Does a changed starting condition produce a meaningful change in the result? These are suggested checks for future testing; this article does not report an experiment.

Simulations should be compared with observations or experiments before they inform consequential decisions. Safety-sensitive applications require independent review and validation. Shared videos should also be clearly identified as generated content whenever viewers could mistake them for real footage.

Who should follow the project?

PAN 2 is especially relevant to researchers working on world models, robotics and agent planning. Access to weights and code allows them to examine failures, adapt the workflow and build repeatable evaluations around their own tasks.

For readers following AI more broadly, the project offers a useful question to keep in mind: how reliably can a model carry an action through a changing scene? Timing, object persistence and causal order all become visible when the answer takes the form of a sequence of frames.

The next useful evidence will come from independent tests and deployment reports. Those should help establish where PAN 2 performs consistently, what tends to break, and how much computation a workable setup requires.