The Institute of Foundation Models (IFM), part of MBZUAI, has released PAN 2, an open world model for physical artificial intelligence. Given a scene and an action described in text, it simulates the outcome as a video. The announcement was made on X on November 10, 2026, and the weights, training code, and inference code are distributed under an Apache 2.0 license.
- IFM, part of MBZUAI, released PAN 2, an open 29-billion-parameter world model that simulates video outcomes from text actions.
- The model pairs an LLM for conceptual reasoning with a diffusion transformer that generates the video.
- IFM’s self-reported results lead open models, but MiniMax-H3 scores higher overall.
- Weights and code are released under Apache 2.0, and IFM recommends a separate safety guardrail.
A world model is a system that predicts how an environment evolves after an action. PAN 2, presented as a general-purpose model, succeeds PAN, released in November 2025. The stated goal is simulation-based reasoning: an agent proposes several plans, PAN 2 simulates each one, and the agent acts according to the best one. For robotics and agent developers, the idea is to test a plan in simulation before executing it.
A Two-Stage Reasoning Process
PAN 2 splits the simulation into two stages. The first, called conceptual, relies on a large language model (LLM), i.e., a system trained to manipulate text. It reads the observation and requested actions, then writes the next world state as structured JSON: events, timestamped actions linked to actors, and camera movement. According to IFM’s report, the default reasoning model used is Kimi K3, but any instruction-following LLM can play this role.
The second stage, perceptual, is handled by a 29B-parameter diffusion transformer, named MMDiT, which predicts the video latents. IFM trained it from scratch on hundreds of millions of annotated videos. A Qwen3.5-9B text encoder, frozen during use, conditions this model, and the Wan 2.2 VAE decoder produces the final images.
IFM argues that the LLM is crucial. Without conceptual reasoning, the full plan completion rate drops from 34.0% to 0.0%, and the step completion rate falls from 75.1% to 8.7%. A simple rewrite into a structured format recovers some of the gap, with 27.8% of full plans completed. IFM also states it discarded video game, animation, and simulation imagery after an initial training phase. A 9B test model trained without them won 58% of blind votes.
Results as Reported by the Publisher
All scores below are self-reported by IFM, which used its own test suite: 6 capabilities evaluated on 300 single-action cases, and 100 multi-step tasks. For full plans, PAN 2 reaches 34.00%, compared to 32.50% for MiniMax-H3, 26.77% for FLUX 3, and 24.49% for Seedance 2.5. For individual steps, it achieves 75.10%, compared to 75.12% for MiniMax-H3. Across the 6 capabilities, it scores 84.48%, compared to 87.20% for MiniMax-H3, which leads in all six categories.
On the Physics-IQ Verified ranking, PAN 2 scores 49.38, leading the open models, behind the closed model FLUX 3 at 51.11. On RBench, it reaches 57.04, compared to 56.16 for MiniMax-H3. In the human preference test, it shows an Elo rating of 1595 with GPT-6 Astra as the reasoner, compared to 1620 for MiniMax-H3.
On the hardware side, each GPU worker loads the full 29B model in BF16, with block offloading as a fallback. IFM does not publish VRAM figures or generation times. The default sequence duration is 241 frames, or about 10 seconds, and the limit is 20 seconds at 704×1248 resolution and 24 frames per second. Descriptions are capped at 1,536 tokens, the units of text processed by the model.
IFM additionally recommends its separate safety guardrail, pan2-guardrail. In the institute’s tests, it reduced unsafe outputs from harmful text-to-video queries from 33.9% to 1.5%, while blocking 11.5% of benign queries.
