PAN 2
modelYour notes
The second open world model from MBZUAI's Institute of Foundation Models (IFM). Given a current observation (an optional image plus text) and actions written in text, it simulates the outcome as a video of up to 20 seconds at 704×1248 and 24 fps. The intended use is simulative reasoning: an agent imagines the result of each candidate plan before it chooses one. It follows the Generative Latent Prediction design that IFM formalized in "Critique of World Models" and used in the first PAN (November 2025). There are two parts:
- Conceptual reasoner: a language model that writes the next world state as a structured, time-grounded caption. The technical report uses Kimi K3 for this role, and the public Physics-IQ leaderboard entry uses Claude Opus 5.5.
- Perceptual reasoner: a 29B latent-diffusion transformer that renders the caption as video. It uses a frozen Qwen3.5-9B text encoder and the Wan 2.2 VAE.
Only the perceptual backbone is IFM's own trained model. It is released in BF16 under Apache 2.0 as a single checkpoint for both text-to-video and image-to-video. The release includes the pretraining and SFT code (a packed dataloader over precomputed latents, FSDP2, context parallelism and Muon), with support in Diffusers and vLLM. The GitHub repository went public on 2026-11-05. The HuggingFace repository was created on October 7 and last updated on November 9. The model card's evaluations are reported as charts without numeric tables:
- six simulation capabilities over 300 single-action cases;
- 100 egocentric multi-step plans;
- public benchmarks: Physics-IQ Verified, PhyGround and RBench;
- a blind human preference study of 1,400 comparisons.
The model card states plainly that about two in three simulated plans contain at least a step that is not carried out. It also lists physical shortcuts, such as pasta appearing in the colander before it is drained, drift across clips in long interactions, and only coarse control of dexterous actions. IFM recommends running the model behind its separate guardrail. In IFM's tests, the guardrail cut the share of harmful text-to-video requests that produced unsafe video from 33.9% to 1.5%, and it blocked 11.5% of benign text-to-video requests. There is no announcement post or arXiv report yet. It sits alongside Google's Genie 4 as an open counterpart, though PAN 2 is offline video simulation rather than real-time interactive play.