✦
Summary
PAN 2 is a general-purpose world model from MBZUAI's Institute of Foundation Models that simulates how the world changes in response to actions. Given an observation, optionally an image plus text, and text-described actions, it outputs a video of up to 20 seconds at 704 by 1248 and 24 fps. It is built for simulative reasoning: an agent imagines the outcomes of candidate plans before acting, which is useful for agents and robotics. The model was released in October to November 2026 and is fully open under Apache 2.0.
Key Features
Simulates action outcomes as video up to 20 seconds at 704 by 1248 and 24 fps
Text-to-video and image-to-video in one checkpoint
Simulative reasoning and multi-step plan simulation for agents and robotics
29B-parameter MMDiT perceptual backbone plus an LLM conceptual reasoner in a Generative Latent Prediction architecture
Structured, time-grounded world-state captions covering events, timed actions and camera motion
Supported in Diffusers and vLLM with day-zero vLLM-Omni support
Training code, configs and inference code all released under Apache 2.0
Safety guardrail (pan2-guardrail) screening requests and outputs
About PAN 2
The architecture pairs a 29B-parameter MMDiT perceptual backbone with an LLM conceptual reasoner in a Generative Latent Prediction design. A conceptual reasoner first rewrites free-text prompts into structured, time-grounded world-state captions covering events, timed actions and camera motion, and the diffusion model then renders the video. Text-to-video and image-to-video run in one checkpoint, supported in Diffusers and vLLM with day-zero vLLM-Omni support, and training code, configs and inference code are all released.
PAN 2 can be run locally via generate.py with the Hugging Face Diffusers layout checkpoint, and it ships with a safety guardrail, pan2-guardrail, that screens requests and outputs. The code is on GitHub at ifm-ai/PAN2 and the weights are on Hugging Face at IFM/PAN2. It is aimed at researchers and developers building agents that plan by simulation rather than by trial and error in the real world.