MBZUAI Open-Sources PAN 2: 29B World Model That Simulates 20 Seconds of Action Outcomes via Video

The Institute of Foundation Models (IFM) at MBZUAI has released PAN 2, an Apache 2.0 licensed world model. Its public description is straightforward: it combines an LLM planner with a 29B parameter Diffusion Transformer, to simulate the potential outcomes of actions as videos, with a maximum duration of 20 seconds per generation.

The term "world model" is often overhyped. In this context, it can be simply understood as a system that takes action conditions as input, and generates corresponding visual results for a period into the future. Robots and physical AI systems operate in the real world where trial and error comes at a high cost, so using generative video to predict action outcomes is a direction with clear practical value. The highlights of PAN 2 lie in its scale and openness: 29 billion parameters, Apache 2.0 license, and full download accessibility, allowing researchers and developers to directly inspect, reproduce, and modify the model.

But do not rush to conclude that 20 seconds of generation means robots are now ready to work reliably. There is still a gap between visually plausible video generation and physical plausibility. Realistic-looking frames do not guarantee that force, collision, material properties, and contact dynamics are all calculated correctly. 20 seconds is also far shorter than the time horizon of many real-world robotic tasks. Whether PAN 2 can work stably in unseen scenarios, on different robot hardware, and under complex interactions remains to be proven by follow-up code releases, benchmark evaluations, and community reproduction efforts.

Another notable point is its technical roadmap: LLMs take charge of planning and semantic understanding, while the Diffusion Transformer is responsible for generating visual outputs. This division of labor explains why the model puts heavy emphasis on action conditioning. For AI practitioners, areas of interest include its interface design, action representation schema, training dataset, and inference cost. A 29B parameter diffusion model is computationally heavy to run, so the real-world deployment barrier will also depend on the official code release and subsequent optimization work.

All currently public information comes from MarkTechPost's reporting and the project page. The original article link is attached below for you to cross-check details: MBZUAI's IFM Releases PAN 2, an Open 29B World Model for Physical AI.

If you are working on embodied intelligence, robot simulation, or video-based world models, PAN 2 is well worth adding to your watchlist. You can start by reviewing what it is capable of generating, then identify where its limitations and errors lie.