MBZUAI Releases PAN 2 Open-Source World Model for Robotics Simulation

The Institute of Foundation Models (IFM) at MBZUAI has introduced PAN 2, a large-scale "world model" aimed at advancing physical intelligence in AI agents. Released on November 10, 2026, the system allows developers to input a scene and a text-based instruction to generate a simulated video of the resulting action. The model, including its weights and training code, has been made available under the Apache 2.0 license to support the robotics and research communities.

Dual-Reasoning Architecture

The system operates using a two-stage process to ensure simulations are both logical and visually accurate. First, a conceptual reasoning layer uses a large language model to translate instructions into structured data, mapping out specific events and actor movements. Second, a perceptual reasoning layer—a 29-billion parameter diffusion transformer—processes this data to render high-definition video at 24 frames per second. This allows the model to visualize up to 20 seconds of physical interaction from a single starting frame.

Performance and Limitations

According to internal benchmarks, the model leads several competitors in executing multi-step plans, achieving a 34% completion rate on complex tasks. While this marks a high point for currently tested models, the results also highlight the ongoing difficulty of long-form simulation; approximately two-thirds of multi-step plans still fail to complete every instruction accurately. Additionally, the model requires significant computing power, necessitating data-center grade GPUs to run the full 29B parameter set effectively.

Open Access for Developers

By providing the inference code and model weights openly, the developers aim to provide a sandbox where autonomous agents can test plans in a virtual environment before attempting them in the physical world. This approach, known as simulative reasoning, is intended to reduce the risks and costs associated with training hardware-dependent AI. The model succeeds the original PAN released in late 2025 and incorporates a training dataset consisting of hundreds of millions of annotated videos.

What to watch next: Look for whether robotics firms adopt this open-source framework to replace proprietary simulation environments in their development pipelines.