PAN 2 weights and inference code now live on Hugging Face under Apache 2.0

LinusEkenstam · x · 2026-11-13

Follow-up with the Hugging Face link for PAN 2. The model card details a general-purpose world model that takes an observation (optional image + text) and text-described actions to simulate up to 20 seconds of outcome video for agent planning. Architecture: an LLM for conceptual reasoning plus a 29B-parameter MMDiT diffusion backbone trained from scratch on hundreds of millions of annotated videos, decoded via the Wan 2.2 VAE. Apache 2.0 licensed.

Related event: MBZUAI Open-Sources PAN 2, a World Model That Simulates Action Outcomes(2 posts)→

Original post →

More from Research

Research channel →