DEV Community

Cover image for PAN 2 World Model vs Wan 2.2: PAN 2 Wins for Agents
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

PAN 2 World Model vs Wan 2.2: PAN 2 Wins for Agents

The PAN 2 world model, released on 10 November 2026 by the Institute of Foundation Models (IFM) at MBZUAI, is the better choice if you are building agents that need to simulate the consequences of an action before taking it. Wan 2.2 is the better choice if you only want open-weights video generation on modest hardware. PAN 2 ships weights, inference code and training code under Apache 2.0, which is the main reason it stands out (IFM press release).

They are not substitutes: Wan 2.2 is a video generator, PAN 2 a video generator wrapped in a reasoning loop, with a much heavier footprint. This article covers what PAN 2 contains, what it costs to run, where it falls short, and how to decide.

TL;DR

  • PAN 2 is an open world model: give it a scene and actions written in words, and it renders a video of what happens next.
  • The release includes weights, full training and inference code, and documentation under Apache 2.0.
  • The perceptual backbone is a 29B-parameter latent diffusion model, so a single 80 GB-class GPU is the realistic floor.
  • It ranked first among open-weight models on the Physics-IQ Verified leaderboard, according to IFM.
  • IFM's model card says about two in three simulated plans contain at least one failed step.
  • Video only, on smaller GPUs: choose Wan 2.2. Agent loops and robotics-style simulation: choose PAN 2 with the guardrails.

What is the PAN 2 world model?

PAN 2 is a general-purpose world model aimed at what IFM calls physical intelligence. You supply a starting scene (an optional image plus text) and describe actions in words. The model produces a video of the likely outcome, constrained by physics, time and space. IFM uses the term simulative reasoning for the idea that an agent imagines the results of candidate actions before it commits to one (IFM press release).

It succeeds PAN (v1), published in November 2025 (arXiv 2511.09057). IFM was set up by MBZUAI in May 2025, with labs in Abu Dhabi, Silicon Valley and Paris, and its portfolio also includes the K2 series of open-source language models and the Jais Arabic model. The Emirates News Agency reported the launch on November 11, with IFM founder Eric Xing framing PAN 2 as the institute's first step toward practical physical intelligence (WAM).

How does PAN 2 work under the hood?

The design follows Generative Latent Prediction, described in the paper "Critique of World Models" (arXiv 2507.05169). It has two parts.

  1. A conceptual reasoning backbone. This is a large language model that writes the next world state as a structured, time-grounded description: events, timed actions bound to specific actors, and camera motion.
  2. A perceptual reasoning backbone. This is a 29B-parameter MMDiT latent diffusion model, trained from scratch by IFM on hundreds of millions of videos annotated in the same structured format.

The released checkpoint handles both text-to-video and image-to-video. It outputs 704x1248 at 24 fps, up to 20 seconds, in BF16. It relies on a frozen Qwen3.5-9B text encoder and the Wan 2.2 VAE (Wan-AI/Wan2.2-TI2V-5B-Diffusers) to decode latents (Hugging Face model card). Notably, the reasoner is swappable. The model card says any instruction-following LLM can be used, and lists full-plan completion in this order: Claude Opus 5.5, GPT-6 Astra (medium effort), Kimi K3, then Qwen 3.8 Flash.

How does the PAN 2 world model perform?

IFM scored the model on six dimensions: physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamics. PAN 2 placed just behind MiniMax-H3 and ahead of four other models. It completed more full plans than any other model in the multi-step test, and it came third overall on the Physics-IQ Verified leaderboard, first among open-weight models (IFM press release), a placement echoed by independent write-ups at Marktechpost and Middle East AI News.

The single-action set has 300 cases (201 image-to-video, 99 text-to-video). The multi-step set has 100 egocentric everyday tasks across 200 prompts and 456 steps, with PAN 2 reported as a mean and standard deviation over three seeds. These are IFM's own results, so treat them as a vendor claim until independent replications appear.

What hardware does PAN 2 need?

The weights are 29B parameters in BF16, which is two bytes per parameter. That gives roughly 58 GB for the weights alone, before activations. This is my own arithmetic from the model card's precision field, not a measured figure. In practice a single 80 GB-class card is the floor for the backbone. The repository's generate.py runs one worker per GPU, and each worker holds the full model, so adding GPUs adds throughput rather than splitting the model.

Wan 2.2's TI2V model in the repository name above is 5B parameters, so it fits far more modest setups. Check its licence terms yourself before commercial use, as I have not covered them here.

How do you deploy PAN 2 yourself?

Steps from the model card and repository README (GitHub):

  1. Download the transformer with hf download IFM/PAN2 --include "transformer/*", then the text encoder with hf download Qwen/Qwen3.5-9B and the VAE with hf download Wan-AI/Wan2.2-TI2V-5B-Diffusers.
  2. Convert the diffusers-layout weights (PAN2Transformer3DModel) once so that the repository's generate.py can load them.
  3. Run bash inference/inference.sh <CKPT_PATH> <OUTPUT_DIR> [PROMPTS_JSON], or submit it with sbatch.
  4. For the full agent workflow, rewrite free-text prompts into structured world-state captions first. The repository includes a prompt-rewriter skill for Codex and Claude Code that applies the caption policy and checks the text-encoder budget (soft limit 1,480 tokens, hard capacity 1,536) (Hugging Face model card).
  5. Collect results from <OUTPUT_DIR>/{t2v,i2v}/*.mp4, each next to a *_config.json.

Training configs (config/pretrain_29B_720p.yaml and config/sft_29B_720p.yaml) ship in the repo, so you can retrain on your own latents. Few world models are released with full training code, which makes this a real differentiator.

What safety guardrails come with it?

The separate pan2-guardrail repository (GitHub) screens requests with Qwen3Guard-Gen-4B before generation and checks output frames with the Q16 classifier afterwards. In IFM's internal tests with Kimi K3 as the reasoner, the share of harmful text-to-video requests producing an unsafe video fell from 33.9% to 1.5%, and the guardrail stopped 11.5% of benign text-to-video and 3.3% of benign image-to-video requests (Hugging Face model card). Budget for that friction if you deploy it.

What are the limits of the PAN 2 world model?

IFM is candid about them on the model card:

  • About two in three simulated plans contain at least one step that is not carried out.
  • Results are sensitive to prompts; some out-of-distribution examples used hand-refined prompts.
  • Appearance drifts across clips, so long-horizon consistency remains an open problem.
  • Text actions are coarse for dexterous manipulation.
  • The simulation can progress while breaking a constraint, such as pasta appearing in a colander before it is drained.

Realistic synthetic video and bias are flagged as risks too. For safety-critical decisions, treat the output as a hypothesis generator, not a forecast.

Which should you pick: PAN 2 or Wan 2.2?

If you need video and nothing else, use Wan 2.2: smaller, easier to host, sufficient for ordinary text-to-video work. If you are building an agent that tests actions in simulation, wants structured captions, needs guardrails or intends to retrain, deploy PAN 2 with the guardrail repo. Of the options in this comparison, it is the only one I found that publishes training code under Apache 2.0. Note that I compared documented specifications and did not run both models side by side.

For related open-weights releases, see Runway's Praxis 1 and its GPU verdict, neural network physics simulation from MIT and UBC, self-hosting LongCat 2.5, SplatAnything for 3D Gaussian splatting and Kimi K3.1 Pro, since Kimi K3 is one of the supported reasoners.

FAQ

Q: Is the PAN 2 world model really open source?

A: The weights, training code, inference code and documentation ship under Apache 2.0, which permits commercial use. The text encoder and VAE come from their own repositories with their own terms.

Q: How much GPU memory does PAN 2 need?

A: The 29B backbone in BF16 is roughly 58 GB of weights before activations, so an 80 GB-class GPU is the realistic minimum (derived from the model card's precision field, not measured).

Q: Can I use a different language model as the reasoner?

A: Yes. The model card states that any instruction-following LLM works, and ranks full-plan completion: Claude Opus 5.5 first, then GPT-6 Astra, Kimi K3 and Qwen 3.8 Flash (Hugging Face model card).

Q: How long are the videos PAN 2 generates?

A: Up to 20 seconds at 704x1248 and 24 fps from one checkpoint that handles both text-to-video and image-to-video.

Q: Does PAN 2 always follow the planned actions?

A: No. IFM reports that about two in three simulated plans contain at least one step that is not carried out, so verify results before relying on them.

Q: Are the guardrails required?

A: Recommended and distributed separately; they sharply cut unsafe outputs in IFM's tests but block some benign requests.

Top comments (0)