Editorial illustration for MBZUAI IFM Releases PAN 2, an Open 29B World Model for Physical AI
PAN 2: Open 29B World Model for Physical AI
The Institute of Foundation Models at MBZUAI put out PAN 2 on November 10, 2026, announcing it on X as an open world model for physical AI. The system takes a scene and a text description of an action, then generates video showing how that action plays out. Weights, training code and inference code are all released under Apache 2.0, a notable choice given how much of this category stays locked behind APIs.
PAN 2 is a 29-billion-parameter diffusion transformer built on IFM's own Generative Latent Prediction architecture, outlined in the institute's Critique of World Models paper. It replaces PAN, which IFM released a year earlier in November 2025. The pitch targets robotics and agent developers directly: instead of letting a robot or software agent act on a plan blind, run that plan through PAN 2 first and watch the simulated outcome before committing to anything in the real world.
That framing, simulative reasoning before action, is where IFM wants PAN 2 to earn its keep. Whether it actually holds up against multi-step plans and competing models is a separate question, and the numbers IFM reported tell a mixed story.
The Institute of Foundation Models (IFM) at MBZUAI released PAN 2, an open world model for physical intelligence. The PAN 2 world model takes a scene and a text action and simulates the outcome as video.
Why this matters
A 29B open model that scores 49.38 on Physics-IQ Verified, ahead of every other open model on that leaderboard, gives robotics teams something they've lacked: a way to pressure-test a plan without burning real hardware cycles or risking a real arm, drone, or warehouse robot. The Apache 2.0 license matters as much as the benchmark numbers. Weights, training code, and inference code all ship together, so researchers can actually inspect how PAN 2 handles edge cases rather than trusting a closed API's word for it.
We'd temper the enthusiasm with the numbers IFM itself published. PAN 2 trails MiniMax-H3 on full-plan completion by a hair and loses by nearly three points on the six-capability aggregate. This is a solid open contender, not a clean sweep against closed systems.
For founders building simulation-first robotics pipelines, that's the real takeaway: PAN 2 is good enough to prototype against today, cheap enough to self-host, and open enough to modify. Whether it holds up outside curated benchmarks is the next thing worth watching.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Larger Model Variants Reproduce and Amplify Human Cross-Scale Correlations
From Archives EarthRecord Benchmark Maps Four Decades of Land Change Across 56 Sites
Common Questions Answered
What is PAN 2 and how does it work as a world model for physical AI?
PAN 2 is a 29-billion-parameter diffusion transformer released by MBZUAI's Institute of Foundation Models that functions as an open world model for physical AI. It takes a scene and a text description of an action as input, then generates video output showing how that action would physically play out in the real world.
Why is the Apache 2.0 license significant for PAN 2 compared to other world models?
The Apache 2.0 license is notable because most world models in this category remain locked behind proprietary APIs, whereas PAN 2 releases its weights, training code, and inference code openly. This allows researchers to inspect how PAN 2 handles edge cases and gives them full transparency into the model's operations rather than treating it as a black box.
What is PAN 2's performance score on Physics-IQ Verified and why does it matter?
PAN 2 scores 49.38 on Physics-IQ Verified, placing it ahead of every other open model on that leaderboard. This strong benchmark performance combined with its open accessibility gives robotics teams a way to pressure-test and validate plans without risking real hardware cycles or physical damage to expensive robots and equipment.
What practical benefits does PAN 2 provide for robotics research and development?
PAN 2 enables robotics teams to simulate and verify physical actions in video form before deploying them on real hardware, eliminating the need to burn expensive hardware cycles or risk damaging actual robotic arms, drones, or warehouse robots. This simulation capability allows researchers to test edge cases and validate plans safely in a virtual environment first.
Further Reading
- PAN 2: A Foundation World Model for Physical Intelligence - Institute of Foundation Models, MBZUAI
- A Tutorial on World Models and Physical AI - arXiv
- World Model for Robot Learning: A Comprehensive Survey - arXiv
- Physical AI: A Survey of Models, Learning, Systems, and Open Challenges - arXiv
- Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI - arXiv
