PAN2: Generate 20-Second Physics Simulation Videos with 29B Parameters

IFM/PAN2

·2026.11.17 01:08

Given the current observation state and a text-described action, it simulates the impact of that action on the world as a video of up to 20 seconds. PAN 2, developed by the Institute of Foundation Models at MBZUAI, supports simulative reasoning, allowing agents to virtually verify outcomes before taking actual actions, rather than simply generating videos.

1 / 4

It operates by combining a 29B-parameter MMDiT-based perceptual backbone with a Qwen3.5-9B text encoder. The language model creates a structured world state description specifying time and actors, and the diffusion model generates a 24fps video at 704x1248 resolution based on this. It supports both text-to-video and image-to-video modes with a single checkpoint.

Unlike existing video generation models, it is specialized in world modeling focused on physical laws and causality. It demonstrates competitiveness in physical consistency evaluations on public benchmarks such as Physics-IQ Verified and PhyGround, and recorded high preference in a blind evaluation involving 13 reviewers. It is released under the Apache 2.0 license, allowing free use for research and development purposes.

It is suitable for robots or autonomous agents to simulate the results of each step before executing complex multi-step plans, enabling the selection of the optimal path. There is a possibility of failure in some steps during long plan simulations, and because it is sensitive to prompt phrasing, using a structured caption format is recommended. For safety, it is recommended to use it together with pan2-guardrail.

HuggingFace
HuggingFace model

IFM/PAN2

The original page has no description.

image-text-to-video

This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.