MBZUAI Opens PAN 2, a 29B World Model that Generates 20‑Second Video Simulations

The Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has made PAN 2 publicly available under an Apache 2.0 licence. According to MarkTechPost, PAN 2 is a world model that couples a large‑language‑model planner with a 29‑billion‑parameter diffusion transformer, allowing it to simulate the outcomes of physical actions for up to twenty seconds of video. The release signals a new level of openness for high‑capacity, physically grounded generative AI.

What is PAN 2 and how does it work?

PAN 2 is a "world model" – a type of AI that learns a representation of how the physical world behaves and can predict future states when given a plan. The model consists of two main components. First, an LLM planner interprets a textual or symbolic instruction and produces a sequence of high‑level actions. Second, a diffusion transformer, sized at 29 billion parameters, renders those actions as visual frames, stitching them together into a short video clip. By chaining the planner’s intent with the diffusion model’s generative capacity, PAN 2 can transform a brief description like “throw a ball and watch it bounce” into a coherent twenty‑second video that respects physics‑like motion.

Why does an open‑source 29‑billion‑parameter world model matter?

Most world‑model research has remained behind corporate firewalls because the compute and data required to train models of this scale are substantial. By releasing PAN 2 under Apache 2.0, IFM provides the community with a fully documented, reproducible checkpoint that can be fine‑tuned, inspected, or integrated into downstream applications without licensing fees. The size of the diffusion transformer (29 B) places it in the same class as many commercial foundation models, yet its focus on simulating physical outcomes distinguishes it from purely text‑or image‑centric systems. Open access encourages academic benchmarking, safety analysis, and creative experimentation, accelerating the collective understanding of how to align generative video with real‑world dynamics.

What practical applications can creators and researchers explore?

Because PAN 2 outputs video directly, it opens a range of use cases:

• Rapid prototyping of physical scenarios – Engineers can describe a mechanical setup and instantly receive a visual simulation, helping to spot design flaws before building hardware.
• Content generation for short‑form media – Creators can feed a script or storyboard to the LLM planner and obtain a ready‑to‑edit video clip, cutting down production time for ads, tutorials, or social posts.
• Training data for downstream models – Researchers can generate diverse, controllable video datasets to train perception or reinforcement‑learning agents, especially in domains where real footage is scarce.
• Interactive storytelling – Game designers can experiment with narrative branches by describing player actions and visualizing the result, informing level design and cinematics.

These applications leverage the model’s ability to render physically plausible motion without requiring a separate physics engine, making it a versatile bridge between language, planning, and visual synthesis.

How can developers access and run PAN 2 today?

The model is hosted on public repositories linked from the IFM release page. Because it is Apache‑2.0, developers can download the checkpoint, the accompanying inference code, and the prompt‑format specifications. The typical workflow involves:

1. Install the required dependencies – PyTorch, a diffusion‑model library compatible with large transformers, and an LLM inference backend (e.g., HuggingFace Transformers).
2. Prepare a planner prompt – A short natural‑language description of the desired action sequence.
3. Run the planner – The LLM component outputs a structured plan (often a list of action tokens).
4. Feed the plan to the diffusion transformer – The model generates video frames sequentially, stitching them into a short clip.
5. Post‑process – Optional up‑scaling, frame‑rate adjustment, or editing with standard video tools.

IFM provides Docker images and cloud‑ready scripts, so users with modest GPU resources (e.g., a single A100) can produce a twenty‑second clip in a few minutes. The open licence also permits commercial deployment, provided attribution is retained.

How does the open licence influence adoption and research?

An Apache 2.0 licence grants users the freedom to modify, distribute, and incorporate the code into proprietary products without paying royalties. This lowers the barrier for startups and larger enterprises to embed PAN 2 into pipelines such as automated video advertising, virtual prototyping, or AI‑augmented game engines. Academics benefit from the ability to publish reproducible results, conduct ablation studies, and explore safety‑critical aspects (e.g., bias in motion generation) without negotiating separate agreements. The transparency also encourages community‑driven improvements—contributors can submit patches that enhance efficiency, add new conditioning modalities, or integrate the model with existing toolchains.

Grafy perspective: trying PAN 2 with Grafy tools

While Grafy does not host PAN 2 directly, the platform’s modular apps make it easy to experiment with the model’s output. Users can upload a generated video clip to Film Studio to add narration, subtitles, or additional visual effects, turning a raw simulation into a polished short film. For creators focused on visual assets, Creator can be used to generate complementary images or textures that match the style of the PAN 2 video, ensuring visual consistency across a project. Developers interested in prompt engineering can track different planner prompts and their resulting videos with Graph, building a branching history of experiments that can be revisited or shared. If you prefer a quick, no‑code exploration, the Chat app lets you ask the AI to explain PAN 2’s architecture or suggest prompt templates, while Reader can ingest the IFM release paper for deeper technical queries.

---

Source: MarkTechPost (https://www.marktechpost.com/2026/11/10/ifm-pan-2-open-world-model/)

Frequently asked questions

How long can PAN 2 simulate video for a single prompt?

PAN 2 is designed to generate up to twenty seconds of video per inference run, which balances visual fidelity with computational cost.

Is the 29‑billion‑parameter diffusion transformer the same size as other commercial models?

Yes, the diffusion transformer component matches the parameter count of many leading commercial foundation models, but it is specialized for video generation rather than static image synthesis.

Can I use PAN 2 for commercial video production without paying royalties?

Because PAN 2 is released under the Apache 2.0 licence, you may incorporate it into commercial workflows without royalty payments, provided you retain the required attribution.

---

Source: MarkTechPost — MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI