MBZUAI Open-Sources PAN 2, a 29B Foundation World Model for Physical Intelligence
On November 11, MBZUAI's Institute of Foundation Models (IFM, ericxing's team) released PAN 2, an open-weight foundation world model built for physical intelligence. The team argues that world models should operate in an "Aristotelian world" rather than a "Platonic" one—instead of predicting the next token like a language model, they should predict how the world evolves after each action, anchoring predictions in time, physics, and space. The core capability is "simulation-based reasoning": letting agents rehearse "what would happen if" in imagination before acting.
Confirmed
- The architecture uses a two-track split: a conceptual layer where an LLM reasons about what happens next and how to achieve goals, and a generative layer—a 29B-parameter diffusion model—that predicts and generates concrete world states.
- The model achieves 75.1% step accuracy in evaluations; demos include the 29B model simulating physical tasks such as a robot folding clothes.
- The team lays out six capability dimensions a world model should have, including adherence to physical laws and historical consistency.
- The vLLM team announced day-zero support for PAN 2, and the model is released with open weights.
Why it matters
- PAN 2 pushes "world models" from concept to a usable open-source foundation model, giving robotics and physical intelligence research a general base for simulation-based reasoning before acting; vLLM's day-zero support also means a low deployment barrier, letting the community quickly reproduce and extend the work.
2026-11-11 ~ 2026-11-11 · 11 related posts
Primary sources
- Open-weight world model PAN 2 launches with Day-0 vLLM support for text/image-to-video — vllm_project ·
- MBZUAI releases open world model PAN 2: LLM planning plus 29B diffusion hits 75.1% step accuracy — rohanpaul_ai ·
- 29B world model PAN 2 lands in Hugging Face Diffusers with t2v and i2v workflows — rohanpaul_ai ·
- [source] Open-weight world model PAN 2 launches with Day-0 vLLM support for text/image-to-video — vllm_project · 2026-11-11
- [source] MBZUAI releases open world model PAN 2: LLM planning plus 29B diffusion hits 75.1% step accuracy — rohanpaul_ai · 2026-11-11
- MBZUAI Releases PAN 2, an Open 29B World Model Simulating Robots Folding Shirts — rohanpaul_ai · 2026-11-11
- MBZUAI releases PAN 2, an open-weights foundation world model for physical intelligence — Scobleizer · 2026-11-11
- IFM releases PAN 2, an open world model that predicts the next world, not word — HongyiWang10 · 2026-11-11
- [source] 29B world model PAN 2 lands in Hugging Face Diffusers with t2v and i2v workflows — rohanpaul_ai · 2026-11-11
- IFM open-sources PAN2: 29B latent diffusion video model, 20s 720p, training code included — Aggressive-Zebra-949 · 2026-11-11
- IFM unveils open foundation world model PAN 2 trained on millions of hours of robot data — ZhitingHu · 2026-11-11
- Eric Xing details PAN 2 world model: GLP architecture and 6 dimensions of physical intelligence — HeyAmit_ · 2026-11-11
- IFM open-sources PAN 2, a foundation world model that predicts the next world state — HeyAmit_ · 2026-11-11
- IFM releases PAN 2, a 29B general-purpose world model simulating actions as 20s videos — pmttyji · 2026-11-11