PAN 2 explained: an open 29B world model, Apache 2.0

By LLMTest Lab · Nov 11, 2026 · 4 min read releasevideoopen-weight
On this page

On this page

  1. What it is
  2. What the vendor numbers say
  3. What it costs to run
  4. Limits IFM states
  5. Who should care
  6. What we will test

Not tested by us yet. PAN 2 is a video model with open weights and no hosted service, and our harness never runs video models, so we have not generated a single clip. Everything below comes from the maker's release as reported by CellCog's technical summary and the launch coverage. We'll update this post when someone publishes independent results or we can run it.

PAN 2 is an Apache 2.0 world model from the Institute of Foundation Models (IFM) at MBZUAI, announced on 10 to 11 November 2026. You give it a scene description, an optional starting image and an action in words, and it returns a video of what happens next. Its own numbers put it level with MiniMax-H3 on multi-step plans, behind it on single actions and human preference, and ahead of Cosmos 3 Super and LTX 2.5. It is for robotics and simulation researchers who can host an 86 GB checkpoint, not for people who want a video tool.

PAN 2
Maker Institute of Foundation Models, MBZUAI
Diffusion model 29.2B parameters, 44 blocks, trained from scratch
Planner Any instruction-following LLM, not included
Output Up to 20 s, 704 x 1248, 24 fps (default about 10 s, 50 sampling steps)
License Apache 2.0 for IFM code and weights; text encoder (Qwen3.5-9B) and decoder (Wan 2.2 VAE) keep their own licenses
Download About 86 GB on Hugging Face
Hosted API, price None listed

What it is

PAN 2 has two parts. A language model writes a structured, time-stamped description of the scene: who acts, when, and how the camera moves. The diffusion model then renders that description as video. Only the diffusion model ships. IFM says any instruction-following LLM can do the planning and tested Claude Opus 5.5, GPT-6 Astra, Kimi K3 and Qwen 3.8 Flash, with Kimi K3 in most studies.

The planner matters a lot by IFM's account. With no language model, no multi-step plan completed. Rewriting the action into the structured format got 27.8% of plans through, and reasoning before writing got 34.0%.

What the vendor numbers say

All of these are IFM's own, with no independent replication yet.

  • Multi-step test (100 everyday first-person tasks, 456 steps, judged by a model): PAN 2 completes 75.10% of steps and 34.00% of full plans. MiniMax-H3 gets 75.12% and 32.50%. FLUX 3 gets 72.44% and 26.77%, Seedance 2.5 gets 68.95% and 24.49%.
  • Single actions (300 cases): PAN 2 scores 84.48% overall against 87.20% for MiniMax-H3, which leads on all six capabilities.
  • Human preference (13 reviewers, 1,400 blind comparisons, open models only): MiniMax-H3 has an Elo of 1620 and a 68.8% win rate. PAN 2 scores 1595 with GPT-6 Astra as planner and 1559 with Claude Opus.
  • Physics-IQ: 49.38 with Claude Opus 5.5 in front, 37.48 in its default setup against 42.91 for MiniMax-H3.

The plan-completion lead over MiniMax-H3 is 1.5 points on a spread of about 3 points between runs. Read it as a tie. IFM's two documents also disagree on the weaker rivals: the blog reports a full-plan rate of 4.0% for LTX 2.5, the model card 13.00%, and neither explains the gap.

What it costs to run

IFM lists one worker per GPU, each holding the full model, plus an offload setting for tight memory. It gives no minimum GPU memory and no time per clip. Prompts must use the structured format, capped at 1,536 tokens per caption, so free text has to pass through the planner first. That is a second model to host or pay for on every request. If you only need a single short clip from a text prompt, our look at how an open robotics release reads on paper is closer to what you want than a world model.

Limits IFM states

About two in three simulated plans contain at least one step that is not carried out. The model is sensitive to wording, drifts in appearance across long scenes, and can reach a requested end state while breaking physics along the way. IFM recommends a separate guardrail: in its tests it cut unsafe output from harmful text requests from 33.9% to 1.5% but also blocked 11.5% of harmless ones. The technical paper was not yet public at launch. CellCog, our main source, uses Seedance 2.5 and Claude models and says it has not run PAN 2.

Who should care

  • Robotics and simulation researchers. Full training and inference code under Apache 2.0 is rare for world models. Compare it on paper with Nvidia's GR00T 2 and Gemini Robotics 2.5, which are action models rather than video predictors.
  • Teams picking a video model for content. Skip it. MiniMax-H3 beats it on IFM's own preference test, and PAN 2 has no hosted API.

What we will test

When we can, we would run 30 first-person tasks through PAN 2 with two different planners and score each step by hand against the written plan, recording GPU memory, seconds per clip and the share of plans that complete. Until then, every quality claim here is a vendor claim.

Related articles

Gemini Robotics 2.5 explained: 30-100% better, partners only
Gemini Robotics 2.5 is DeepMind's newest robot-control model. What it reports against version 2, who can use it today, and why the numbers are vendor-only.
GR00T 2 explained: Nvidia's open world action model for robots
GR00T 2 is Nvidia's open world action model for humanoid robots. What Nvidia reports, what is still missing, and who should look at it first.