MBZUAI's IFM releases PAN 2, an open 29B-parameter world model that simulates the outcome of actions
A language model writes a time-grounded description of what happens next and a 29B diffusion model trained from scratch renders it as video of up to 20 seconds; weights, training and inference code are under Apache 2.0.
Why it matters
World models that let an agent rehearse candidate actions before acting are a central bet for robotics and physical AI, and most strong video models are closed. PAN 2 splits the job: a multimodal language model handles what happens and when, while the diffusion model supplies the visual detail, so each prediction can be checked against real footage. It ships with a Qwen3.5-9B text encoder and the Wan 2.2 VAE and runs in Hugging Face Diffusers and vLLM. Watch for the full paper IFM says is coming and for independent tests of its planning claims.
Key facts
- Developed by the Institute of Foundation Models (IFM) at MBZUAI
- 29B-parameter MMDiT trained from scratch on hundreds of millions of videos
- Video output 704x1248 at 24 fps, up to 20 seconds
- Text encoder Qwen3.5-9B; VAE from Wan 2.2
- Weights, training and inference code under Apache 2.0
- Supported in Hugging Face Diffusers and vLLM