With PAN 2, the Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) released an open PAN 2 world model on 11 November 2026. The idea: you give the model a starting scene and an action in words, and PAN 2 simulates what happens next as a video of up to 20 seconds. Weights, training code and inference code are available under the Apache 2.0 license, which is rare for world models of this size. The benchmarks, however, come mostly from IFM itself, and the list of documented weaknesses is long.
What the PAN 2 world model is meant to do
A language model predicts the next word; a world model predicts the next state of the world. That is how IFM puts it in its PAN 2 technical blog. The goal is what the team calls simulative reasoning: an AI agent first plays through possible actions in simulation, compares the outcomes and then picks the best option – without costly, slow or irreversible trials in the real world. That matters most in robotics, where every failed attempt costs time, material or hardware.
As input, PAN 2 takes a current observation – optionally an image plus text – and an action described in text, including its duration. The output is a video of the consequences. According to the model card on Hugging Face, the model renders at 704 × 1248 pixels and 24 frames per second, for up to 481 frames, or 20 seconds. Simulations can be chained: the last frame of one clip becomes the starting frame of the next.
Architecture: a language model plus video diffusion
PAN 2 follows the Generative Latent Prediction (GLP) architecture developed at IFM and splits the work into two parts:
- Conceptual reasoning: a pretrained large language model describes the next world state as a structured, time-stamped text – who acts when, what happens, how the camera moves.
- Perceptual reasoning: a 29.2-billion-parameter diffusion model (MMDiT), trained from scratch by IFM, fills in the details that language leaves open: contact, deformation, motion. According to the model card, it was trained on “hundreds of millions” of videos annotated in the same text format.
For text encoding and image decoding, PAN 2 relies on third-party components: Qwen3.5-9B as a frozen text encoder and the Wan 2.2 VAE. Notably, the language model used for conceptual reasoning is interchangeable. IFM used different models depending on the test – by default Kimi K3, according to the model card, and Claude Opus 5.5 and GPT-6 Astra for individual benchmarks. Results therefore depend not only on PAN 2 itself but also on which language model writes the scene description.
Benchmarks: strong on plans, but self-reported
IFM built its own test cases for six “dimensions of physical intelligence”: physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamism. On top of that come 100 everyday multi-step tasks (200 prompts, 456 individual steps). The videos are scored by a vision-language model answering yes/no questions; camera motion is verified using estimated geometry.
The key vendor figures:
- PAN 2 completes 75.1 percent of individual steps and 34.0 percent of full plans. According to IFM, that is the highest full-plan completion rate among the models studied.
- On single actions across the six dimensions, PAN 2 places just behind MiniMax-H3 and ahead of the four other comparison models, according to the press release.
- On the public Physics-IQ Verified benchmark, PAN 2 would rank first among open models and third overall in the leaderboard setting.
- In a blind comparison with 140 prompts, 13 reviewers and 1,400 pairwise comparisons, PAN 2 ranks second among four open models.
The comparison covered the open models Cosmos 3 Super, LTX-2.5 and MiniMax-H3 and the closed models FLUX 3 and Seedance 2.5. According to IFM, the closed models were generated only once. There are no independent measurements yet; IFM also produced the Physics-IQ score in its own runs. The numbers should therefore be read as self-reported.
The weaknesses are in the model card
IFM is unusually open about the limits. The model card lists, among other things:
- Long plans often fail: about two in three simulated plans contain at least one step that is not carried out. How faithfully PAN 2 simulates plans that should fail has not yet been measured.
- Prompt sensitivity: results depend heavily on wording; examples beyond the training data and the agent case studies relied on hand-refined prompts.
- Long-horizon drift: each clip starts from at most one frame plus text, so appearance drifts across clips. A world that stays consistent while an agent explores it is only announced for a future post.
- Shortcuts: simulations can violate physical constraints, such as pasta appearing in the colander before it has been drained.
- Coarse actions: text works for “open a drawer”, but specifies dexterous detail such as where each finger touches an object only coarsely.
On top of that comes the misuse risk that any realistic video model carries. IFM recommends running PAN 2 behind a separate safety filter (“pan2-guardrail”) that screens both requests and generated videos. In internal tests, the filter reduced the share of harmful text-to-video requests that yield an unsafe video from 33.9 to 1.5 percent – but it also blocked 11.5 percent of benign text-to-video requests.
An agent learns something it cannot see
The most telling example in the blog: a ball sits in a tube closed by two caps. The blue one comes off, the red one is glued shut – a fact stated only in the text given to PAN 2, not in the image and not to the agent. An agent built on Kimi K3 proposes a plan, PAN 2 simulates it, and the agent judges the result. Only on its sixth attempt does it try the blue cap, and the ball comes out. Along the way, the agent misreads one simulation and believes the red cap has come off. That shows both the potential of the approach and how error-prone the chain of language model, simulation and judgement still is.
Who PAN 2 matters for
PAN 2 is not a product for end users. It is a research tool for robotics, agents and physical reasoning in video. Running it yourself requires serious hardware: according to the GitHub repository, each inference worker holds the full 29-billion-parameter diffusion model per GPU. Diffusers and vLLM are supported. The Apache 2.0 license covers IFM’s own code and weights; Wan 2.2 and Qwen3.5-9B are subject to their own licenses.
PAN 2 is most interesting as an open reference. World models are becoming more important in robotics, and the strongest systems are often closed. With full training code, it becomes possible to verify what such a model can do and where it fails. For comparison with classic language and image models, see our pieces on Meta’s Muse Ember and the image model Seedream 5.5. We recently looked at how vendors present their own benchmarks in our article on Claude Fable 5.2.
