PAN 2 wants robots to think before they move. The Institute of Foundation Models at MBZUAI released the open world model for physical intelligence on November 10, 2026, giving robotics and agent developers a system that takes a scene and a text action, then simulates the outcome as video.
The pitch is direct: test a plan in simulation before acting on it. An agent proposes several plans, PAN 2 simulates each one, and the agent acts on the best option—assuming the simulation gets the important parts right, which is doing a lot of work in that sentence.
Weights, training code, and inference code ship under Apache 2.0. PAN 2 follows PAN, the predecessor released in November 2025, and targets “simulative reasoning” rather than video generation as a standalone visual trick.
A planner, a video model, and a large hardware bill
PAN 2 is a dense 29B-parameter diffusion transformer, or MMDiT. It accepts text plus an optional first frame and produces 704×1248 video at 24 fps for up to 20 seconds, running on data-center GPUs.
Each worker holds the full 29B model in BF16, with block offloading available as a fallback. IFM publishes no VRAM figure or generation time, so the model is open in licensing terms but not exactly casual about hardware.
The system splits simulation into conceptual and perceptual reasoning. For conceptual reasoning, an LLM reads the observation and requested actions, then writes the next world state as structured JSON containing events, timed actions bound to actors, and camera motion.
IFM’s report uses Kimi K3 by default for that step. For perceptual reasoning, the 29B MMDiT predicts video latents, a frozen Qwen3.5-9B text encoder conditions the model, and the Wan 2.2 VAE decoder renders the final frames.
IFM trained the MMDiT from scratch on hundreds of millions of annotated videos. The training process also included gameplay, animation, and simulation footage after early training, and a 9B test model trained without that material won 58% of blinded votes.
The benchmark story is promising—and unfinished
IFM’s own test suite covers six capabilities across 300 single-action cases and 100 multi-step tasks. All numbers are self-reported by IFM, and they show PAN 2 leading the tested models on full plans while losing ground on several broader measures.
PAN 2 reached 34.00% full-plan completion, compared with 32.50% for MiniMax-H3, 26.77% for FLUX 3, and 24.49% for Seedance 2.5. That is the best result among the tested models, but it also means about two in three plans miss a step.
Step completion reached 75.10%, level with MiniMax-H3 at 75.12%. Across all six capabilities, PAN 2 scored 84.48%, behind MiniMax-H3 at 87.20%, which led in every category.
Physics-IQ Verified reached 49.38 in the leaderboard setting, placing PAN 2 first among open models, while closed FLUX 3 scored 51.11. On RBench, PAN 2 scored 57.04, ahead of MiniMax-H3 at 56.16 and Cosmos 3 Super at 50.64.
Human preference tells a less flattering story. PAN 2 recorded a Human Elo of 1595 with a GPT-6 Astra reasoner, behind MiniMax-H3 at 1620, and its Physics-IQ system default was 37.48 versus 42.91 for MiniMax-H3.
The ablation results explain why the planning layer matters. Removing conceptual reasoning dropped full-plan completion from 34.0% to 0.0% and step completion from 75.1% to 8.7%; a format-only rewrite recovered part of the loss, reaching 27.8% full-plan completion.
Developers can run PAN 2 through Hugging Face Diffusers, vLLM, FastVideo, or the repository’s generate.py. The default sampler uses 50 steps, flow shift 9, guidance 2.5, and a 241-frame clip lasting about 10 seconds.
Prompts work best as structured captions with a hard budget of 1,536 tokens. IFM also ships a Conceptual Reasoner skill for Claude Code and Codex, designed to rewrite free text before the model processes it.
Safety comes with its own add-on. IFM strongly recommends pan2-guardrail, which cut unsafe outputs from harmful text-to-video requests from 33.9% to 1.5% in IFM tests, while blocking 11.5% of benign text-to-video requests.
PAN 2’s result is not a solved robotics brain. It is an open attempt to connect language-based planning with visual prediction, giving agents a way to inspect possible futures before committing to one—and giving developers a measurable place to start finding where those futures go wrong.
Based on
- MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI — marktechpost.com
- Robbyant Releases LingBot-Touch to Bring General-Purpose Touch Perception to Robots | AFP.com — afp.com
- DeepMind reveals Gemini Robotics 2.5 and open safety benchmark “ASIMOV-jailbreak” | Ars OpenForum — arstechnica.com
- Humanoid robots are getting more lifelike — cnbc.com
- Are We Ready for Robots That Code Their Own Skills? — spectrum.ieee.org
