MBZUAI releases PAN 2: an open 29B world model that simulates physical actions as video.

The Institute of Foundation Models at MBZUAI has released PAN 2, a 29-billion-parameter open-source world model designed for physical AI and robotics applications. This model takes a scene description and a proposed text action, then generates a video simulation showing the predicted outcome. It's built on a diffusion transformer architecture and outputs videos up to 20 seconds long at 704×1248 resolution. The entire model weights, training code, and inference code are available under the permissive Apache 2.0 license.

For developers building robotics and AI agents, PAN 2 offers a valuable tool to test plans in simulation before executing them in the real world. It reportedly leads six other tested models on multi-step planning tasks, achieving about 34% success on complex plans. However, the model still fails on about two-thirds of longer plans and requires significant data-center GPU resources to run. This represents a meaningful step toward more reliable simulation for physical AI systems.