Mbzuai's ifm unveils pan 2, an open-source 29b world model for physical AI simulation

The Institute of Foundation Models (IFM) at MBZUAI has released PAN 2, an open-weights world model designed for physical intelligence. Announced on X on November 10, 2026, this model allows users to simulate the outcome of a text action within a given scene, generating video outputs. The model's weights, training code, and inference code are available under the Apache 2.0 license. PAN 2, which builds upon the Generative Latent Prediction (GLP) architecture, aims to facilitate simulative reasoning, enabling agents to test proposed plans in simulation before executing them in reality. IFM's testing demonstrated the importance of conceptual reasoning, showing that without it, full-plan completion dropped from 34.0% to 0.0%. The system supports both text-to-video and image-to-video generation and can be run via Hugging Face Diffusers, vLLM, or FastVideo. Furthermore, IFM strongly advises using the separate pan2-guardrail, which reduced unsafe outputs from harmful text-to-video requests from 33.9% down to 1.5%.