MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI
The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has released PAN 2, an open-weight world model aimed at physical AI - systems that need to understand and predict how actions play out in the real world. The model is available under the Apache 2.0 license, meaning developers and researchers can use, modify, and build on it without the restrictions common to more commercially guarded releases.
PAN 2 uses a two-part architecture: an LLM-based planner that reasons about what actions to take, coupled with a 29-billion-parameter diffusion transformer that renders the predicted outcomes of those actions as video. This design allows the model to synthesize up to 20 seconds of simulated footage showing what would happen given a particular action or sequence of instructions - a capability central to training and evaluating robotic and embodied AI systems.
World models have become an important area of focus in physical AI research because they let systems rehearse and evaluate decisions in simulation rather than relying solely on costly real-world interaction. By generating plausible video of action consequences, a world model can serve as a kind of internal simulator for a robot or autonomous agent, helping it plan more effectively. The 29B parameter scale of PAN 2's diffusion transformer puts it among the larger openly available models in this category, which has historically been dominated by proprietary systems.
The release from MBZUAI continues a broader trend of academic and research institutions contributing large-scale open models to the field. With its Apache 2.0 licensing, PAN 2 is positioned as a foundation that the robotics and embodied AI research community can build on directly - whether for simulation, data generation, or as a backbone for downstream physical AI applications. Full details on the architecture, training data, and benchmarks are available via the original release.
