On 11 November 2026, the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) announced PAN 2, an open world model that generates video of how a scene is likely to change when given a written action, with weights and code released under the Apache 2.0 licence. For labs that work on simulation, robotics and embodied AI, it is a freely usable baseline to test, and IFM says itself that long-horizon consistency is unsolved.
What was announced
The Emirates News Agency (WAM) and Middle East AI News both reported the release on 11 November 2026. According to the Middle East AI News report, the package covers the video model weights, training and inference code, a technical blog and documentation. The weights are hosted on the IFM/PAN2 model card on Hugging Face, which describes PAN 2 as a latent diffusion video world model that takes an image and text and simulates how the scene changes given an action.
The model card lists a 29-billion-parameter diffusion backbone in a single checkpoint that handles text-to-video and image-to-video, with output at 704 by 1248 pixels, 24 frames per second and up to 20 seconds. It does not state a release date. Its citation block carries an October 2026 date, so the files may have been visible before the 11 November announcement; the announcement date is the one the news reports give.
PAN 2 follows the first PAN model from the same institute. IFM’s research group also published the WR-Arena benchmark for world models on arXiv in March 2026, which tests action simulation fidelity, long-horizon forecasting and simulative reasoning.
How it works, as described by IFM
Middle East AI News reports that a large language model first writes a time-stamped description of the events that should follow an action, and a video diffusion model trained from scratch on physically dynamic real-world footage then renders it. IFM calls this approach simulative reasoning. On results, the same report says PAN 2 ranks third overall on the Physics-IQ Verified leaderboard and first among open-weight models. These are the developer’s reported figures; independent replication has not been reported in the sources above.
The licence terms
The model card states that PAN 2’s own code and weights are under the Apache License 2.0. Apache 2.0 is a permissive licence, so reuse, modification and redistribution are allowed under its conditions. The card also says third-party components keep their own licences: Wan 2.2 (the VAE), Qwen3.5-9B (the frozen text encoder) and Diffusers. The text encoder and VAE are downloaded from their own repositories, so a lab planning to redistribute a derived system should check each component’s terms separately. The card sets out-of-scope uses: deepfakes or impersonation, misinformation, sexual content involving minors or non-consensual intimate content, infringement of intellectual property or privacy, and safety-critical decisions without independent human review.
Release of weights and code is not the same as release of the training data. See the open weights model entry and the open-source model criteria entry for how the terms differ, and the software license agreement entry for the contractual side of institutional reuse.
What it means for simulation, robotics and embodied-AI labs
The model card names agent and robotics action simulation, physical and causal reasoning in video, evaluation, education and creative work as intended uses. For a lab, an open model has practical value that a hosted service lacks: the checkpoint can be pinned to a version, run on local hardware, and re-run for a paper. That supports computational reproducibility in a way an API that changes under the user does not.
The cost is hardware and setup. The card says each GPU worker holds the full 29B model and lists a specific software stack (Python 3.14, PyTorch 2.12.1, CUDA 13.2 builds, flash-attn-4 and quack-kernels). It does not give a VRAM figure, so a research office budgeting compute should confirm requirements with the lab before committing.
Limits that IFM states
The model card is direct about weaknesses. It says about two in three simulated plans contain at least one step that is not carried out, results are sensitive to prompt wording, appearance drifts across clips in long interactions, and text describes coarse actions well but dexterous, fine-grained motion poorly. It also says the model can skip physical constraints, such as showing an intermediate state out of order. IFM calls long-term world consistency an open challenge and describes the research as early-stage. Outputs can look like real footage, and the developers recommend running the model behind the separately distributed pan2-guardrail, which they say reduces but does not eliminate harmful outputs. A video generator of this kind is not a validated physics simulator, so work that relies on its outputs as evidence about the physical world needs its own validation.
Related coverage on CASRAI
Other recent items cover related ground: MIT and UBC’s WorldParticle single-model simulator and DeepMind’s Gemini Robotics 2.5.
Sources
- WAM: MBZUAI launches open-source PAN 2 model to simulate physical events, 11 November 2026.
- Middle East AI News: MBZUAI’s IFM opens PAN 2 world model, 11 November 2026.
- Hugging Face: IFM/PAN2 model card, checked 2026-11-14.
- arXiv 2603.25887: World Reasoning Arena (PAN Team, IFM), 26 March 2026.
