MBZUAI Releases PAN 2: Open 29B World Model for Physical AI

MBZUAI's Institute of Foundation Models has unveiled PAN 2, an open-source world model designed to simulate physical environments for robotics and embodied AI systems. Officially released on November 10, 2026, the 29-billion parameter model combines language-based planning with video generation to predict action outcomes up to 20 seconds into the future. Released under the permissive Apache 2.0 license, PAN 2 marks a significant contribution to accessible physical AI infrastructure.

Architecture and Technical Design

PAN 2 employs a two-stage architecture that separates high-level reasoning from physics simulation. An LLM planner first interprets task requirements and generates action sequences in natural language. These plans then feed into a 29-billion parameter diffusion transformer that renders predicted outcomes as coherent video sequences. This division of labor allows the system to handle both abstract goal decomposition and detailed physical dynamics within a unified framework.

The diffusion transformer component uses a latent space representation to efficiently model complex physical interactions. By operating in compressed video space rather than raw pixels, PAN 2 achieves 20-second prediction horizons without prohibitive computational costs. The model can simulate object manipulation, navigation scenarios, and multi-step task execution with temporal consistency across the generated sequences.

Release Date and Availability

Officially launched on November 10, 2026, PAN 2 is immediately available for commercial and research use through the Apache 2.0 license. MBZUAI has released model weights, inference code, and technical documentation to facilitate adoption. The institute provides both cloud-hosted API access and downloadable checkpoints for local deployment, enabling researchers and developers to integrate the model into robotics platforms without licensing restrictions.

Applications in Physical AI

PAN 2 addresses a critical gap in embodied AI systems: the ability to mentally simulate action consequences before committing to physical execution. Robotics applications include manipulation planning, where systems can visualize grasping outcomes across different approaches, and navigation in dynamic environments where predictive modeling reduces collision risks. The model's 20-second horizon covers typical task segments in warehouse automation, household robotics, and agricultural systems.

Beyond robotics, the architecture enables synthetic data generation for training other AI systems. Developers can use PAN 2 to create diverse training scenarios without physical prototyping, accelerating development cycles for embodied agents. The model's open licensing removes barriers for academic labs and startups working on next-generation physical AI products.

Open Source Strategy and Model Scale

MBZUAI's decision to release PAN 2 under Apache 2.0 contrasts with proprietary approaches from commercial labs. The 29-billion parameter scale balances capability with practical deployability, fitting within memory constraints of modern GPU configurations while maintaining prediction quality. The institute has documented training procedures and architectural choices, enabling the research community to build derivative models and extend the system to new domains.

• 29 billion parameters optimized for physical simulation tasks
• 20-second prediction horizon for action outcome modeling
• Apache 2.0 license enabling unrestricted commercial use
• Dual-component architecture separating planning from physics rendering
• Supports both cloud API and local deployment configurations

What This Means

PAN 2 lowers entry barriers for physical AI development by providing production-ready world modeling capabilities without licensing costs or usage restrictions. The combination of language-driven planning and video-based physics simulation offers robotics engineers a practical tool for safe exploration of action spaces. As embodied AI systems move from research labs to real-world deployments, open models like PAN 2 accelerate innovation by enabling rapid prototyping and domain adaptation across diverse physical environments.