MBZUAI Releases PAN 2 Open Source World Model

MBZUAI's Foundation Models Institute released PAN 2 on November 10, 2026. It is an open AI model designed to simulate the physical consequences of actions. The system predicts what happens in a real environment after a specific sequence of movements, serving as a foundation to train robots and software agents.

What was announced

PAN 2 is a world model, meaning it simulates the rules of an environment, and it has 29 billion parameters. The project is open source under the Apache 2.0 license, with weights and code available on Hugging Face and GitHub. The model generates video at 24 frames per second, with a resolution of 704x1248 pixels and a default duration of about 10 seconds, up to a maximum of 20. Adding a text planner improves results: without it, only 8.7% of steps are correct, while the full system reaches 75.1% accuracy and 34% completed plans.

How it works

The architecture splits the work into two distinct phases. A first language model analyzes the initial scene and requested actions, writing a structured text description detailing who does what and when. A second model, trained from scratch on hundreds of millions of videos, turns this description into a realistic video using a 9-billion parameter text encoder and a visual decoder. It is like a director writing a detailed script first, and a second team handling only the final footage, ensuring objects keep their shape and position even when the camera moves.

What changes for businesses

For businesses developing physical automation, this tool reduces real-world experimentation costs. For example, a company building robotic arms for warehouses can use the model to simulate lifting boxes of different weights, checking if the grip holds before programming the physical machine. An engineering firm can also virtually test how people move in a new office layout, generating realistic videos of movement to evaluate workflows without building physical mockups.

Limitations and what to watch out for

The model is not interactive and does not run in real time, so it cannot replace a graphics engine for live applications. It requires heavy hardware resources: the model weights alone take up about 58 gigabytes, plus 19 gigabytes for the text encoder. Training data focuses on everyday physical scenes and robotics, so results might not suit stylized art generation without extra training. Sources do not specify if the model already complies with specific European AI regulations, but they recommend using a dedicated safety filter.

How to try it

Code and weights are downloadable from Hugging Face and GitHub. Running it requires an inference script included in the repository, which also supports partial loading from system memory if hardware is insufficient. Developers can use open inference stacks like vLLM-Omni, which ensures immediate support for text-to-video and image-to-video generation.