MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI

What happened

On November 10, 2026, the Institute of Foundation Models at MBZUAI announced the release of PAN 2, a new open world model designed for simulating physical environments. PAN 2 generates video predictions of scenes based on a given description of actions, supporting applications in robotics and agent planning. The model and its training and inference code are available under an open-source license. Developers can use PAN 2 to test plans in simulation before executing them in the real world.

Key details

• PAN 2 has 29 billion parameters and uses a diffusion transformer architecture.
• The model outputs videos up to 20 seconds long at a resolution of 704×1248 and 24 frames per second.
• It leads tested models in multi-step plan completion with a 34.0% success rate on 100 tasks (self-reported).
• PAN 2 requires data-center GPUs to run, with each worker loading the full model in BF16 precision.
• The model, code, and weights are released under the Apache 2.0 license.

Why it matters

Robotics and AI developers now have access to a large, open-source model for simulating physical actions and plans, which can help reduce the risk and cost of real-world testing. However, the high hardware requirements may limit use to organizations with significant computing resources. The model’s strong performance on multi-step planning tasks offers a new tool for those building advanced agents and robotics systems.