11 November 2026 · 3 min read
PAN 2: MBZUAI’s IFM Releases Open 29B World Model for Physical AI
IFM at MBZUAI released PAN 2, an open 29B-parameter world model that simulates video outcomes from text actions. Weights and code are under Apache 2.0.
On 10 November 2026, the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) released PAN 2, an open world model designed for physical intelligence [1][3]. The model takes a starting scene and a text action, then simulates the outcome as a video. Weights, training code and inference code are available under the Apache 2.0 licence [1][2].
What PAN 2 does
PAN 2 is a general-purpose world model that implements what IFM calls simulative reasoning. An agent proposes a plan, PAN 2 simulates how the scene would change, and the agent can then choose the best course of action without real-world trial and error [3]. The model outputs video at 704×1248 resolution, 24 frames per second, up to 20 seconds [1][2].
How it works
The system splits reasoning into two parts [1][3]:
- Conceptual reasoning: A large language model (LLM) reads the observation and requested actions, then writes a structured, time-grounded description of the next world state as JSON. Any instruction-following LLM can fill this role; IFM’s report uses Kimi K3 by default.
- Perceptual reasoning: A 29B-parameter multi-modal diffusion transformer (MMDiT), trained from scratch on hundreds of millions of annotated videos, predicts the video latents. A frozen Qwen3.5-9B text encoder conditions the generation, and a Wan 2.2 VAE decoder renders the final frames.
An ablation study showed that removing the LLM planner drops full-plan completion from 34.0% to 0.0% [1][2].
Performance
All performance figures are self-reported by IFM, measured on its own test suite covering 300 single-action cases and 100 multi-step tasks [1][2].
- Full-plan completion: 34.00%, the highest among six tested models. MiniMax-H3 scored 32.50%, FLUX 3 26.77%.
- Step completion: 75.10%, essentially level with MiniMax-H3 at 75.12%.
- Six core capabilities (physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance, objective dynamics): 84.48% overall, behind MiniMax-H3’s 87.20%.
- Physics-IQ Verified (leaderboard setting): 49.38, first among open-weight models. Closed model FLUX 3 scored 51.11.
- RBench: 57.04, slightly ahead of MiniMax-H3 (56.16) and Cosmos 3 Super (50.64).
- Human Elo: 1595, behind MiniMax-H3 at 1620.
About two in three simulated plans still miss at least one step, a limitation the team acknowledges openly [1][2].
Availability and running the model
PAN 2 weights are on Hugging Face (repository created 7 October 2026) and the code is on GitHub (since 5 November 2026) [4]. The model can be run via Hugging Face Diffusers, vLLM, FastVideo or the repo’s generate.py script [1][2]. vLLM-Omni merged Day-0 support for PAN 2 on 10 November, before the official announcement [4].
Each GPU worker loads the full 29B model in BF16; block offloading is available as a fallback. IFM has not published specific VRAM requirements or generation times [1][2]. There is no hosted API or pricing.
Prompts work best as structured captions. The repository includes a prompt-rewriter skill that enforces a token budget of 1,536 tokens [1][2]. A separate guardrail model, pan2-guardrail, is recommended to filter unsafe outputs [1][2].
What this means for developers
For robotics and agent builders, PAN 2 offers a fully open, trainable world model that can simulate multi-step plans before execution. The Apache 2.0 licence and release of full training code lower the barrier for research and commercial adaptation [3]. The model’s reliance on an LLM for conceptual reasoning means developers can swap in different planners to suit their domain.
Hardware demands are high, and long-horizon plan completion remains unreliable. Still, the release provides a foundation for simulative reasoning in physical AI, with the transparency that comes from open weights and code.
Sources
- MBZUAI's IFM Releases PAN 2, an Open 29B World Model for Physical AI
- The Institute of Foundation Models unveils PAN 2, an open world model for physical intelligence
- PAN 2 Shipped Quietly, Then Got a Serving Path First
- world-model
- physical-ai
- open-source
- video-generation
- robotics
Researched and drafted with AI from the sources above.
Learn to build with AI
Hands-on tutorials on Cursor, Claude Code, AI agents and more.