PAN 2 Open World Model Released With Training Code

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence has released PAN 2, an open world model that takes a starting scene and a written action and simulates a video of what happens next. In a 10 November 2026 press release, IFM describes the system this way: “Give it a starting scene and an action in words, and PAN 2 simulates a video of what happens next, grounded in physics, time, and space.” The release includes model weights, full training and inference code, a technical blog and documentation, all under the Apache 2.0 licence.

That is a genuine opening of the stack, and it is also a narrower claim than a general, independently proved simulator of the physical world. The official model card is frank about one limit in particular: “About two in three simulated plans contain at least one step that is not carried out.” The rankings IFM cites for PAN 2 are the institute’s own results.

PAN 2 turns a scene and a text action into up to 20 seconds of video

The Hugging Face card describes a two-part system. A large language model acts as a conceptual reasoner: it reads the observation and the requested actions and writes the next world state as a structured, time-grounded description. A 29-billion-parameter MMDiT diffusion backbone, trained from scratch, then turns that description into video. Output is 704×1248 at 24 frames a second, for up to 20 seconds. The frozen text encoder is Qwen3.5-9B and the decoder is the Wan 2.2 VAE; both are downloaded from their own repositories and used under their own licences, not as part of IFM’s Apache 2.0 grant.

IFM’s founder and MBZUAI president, Eric Xing, said: “Physical intelligence begins when a system can simulate what its actions would do. PAN 2 is our foundation for that.” Hector Liu, director of IFM’s Silicon Valley lab, said the open release “means releasing the model weights alongside the training and inferencing code.” Those quotes are from the 10 November press release. The same laboratory already appears in TSN’s coverage of Jais 3, IFM and Inception42’s open Arabic models. PAN 2 is a different product: a world model, not a language model.

The card also recommends running PAN 2 behind a separate pan2-guardrail. In IFM’s internal tests, with Kimi K3 as the conceptual reasoner, that guardrail reduced the share of harmful text-to-video requests that produced an unsafe video from 33.9% to 1.5%. It also stopped 11.5% of benign text-to-video requests. IFM says the guardrails reduce, without eliminating, the risk of harmful outputs.

IFM’s own rankings are not an independent verdict on world simulation

IFM’s press release places PAN 2 “just behind MiniMax-H3 and ahead of four other models” on six core dimensions, says it “completed more full plans than any other model tested” in multi-step planning, and ranks it third overall on Physics-IQ Verified, and first among open-weight models. Those are developer claims. They are not a third-party audit, and they are not a general proof that the model reliably simulates the world.

The same official card that hosts the work also lists the gaps. Besides the two-in-three figure, IFM says the quantitative study covers open-loop plans designed to succeed, and that “how faithfully PAN 2 simulates plans that should fail is not yet measured.” The model is prompt-sensitive; examples beyond the training data and the agent case studies relied on hand-refined prompts. Appearance drifts across clips in long interactions, because each clip is simulated from at most one starting frame plus text. Text actions are too coarse for dexterous contact. And PAN 2 “may let a simulation progress while violating physical constraints,” for example pasta appearing in a colander before it is drained.

That last point is the one that should sit beside any Physics-IQ headline. TSN has already had to make a similar distinction for Odyssey-3, whose headline physics score was a best-of-eight and whose every benchmark was Odyssey’s own. An open-weight world model is still a welcome artefact. It does not, by itself, settle how often the simulated plan is the plan that would actually happen.

The wider field is moving quickly, from DeepMind’s Genie 4, which adds sound to worlds that last a few minutes to Tencent Hunyuan’s open-sourced Hy-Nomas interactive world model. Those systems are useful context. They are not a substitute for reading PAN 2’s own limitations.

The next check is whether anyone repeats the Physics-IQ and planning numbers under a protocol that is not IFM’s, and whether later cards narrow that two-in-three miss rate. Until then, the release to take seriously is the one that is actually on the table: Apache 2.0 weights and training code, a 20-second simulator, and a documented habit of skipping steps.

Sources

IFM, “PAN 2” press release, 10 November 2026. https://ifm.ai/pan2/press-release/
Hugging Face, IFM/PAN2 model card. https://huggingface.co/IFM/PAN2
GitHub, ifm-ai/PAN2, Apache 2.0 training and inference code. https://github.com/ifm-ai/PAN2