An open world model renders 20 seconds of predicted action as video

single source· 1 articles · confidence: low · first seen 2026-11-11 07:38 UTC

What this means for you

Nothing to act on yet. The permissive licence means a team could build on PAN 2 without negotiating terms, but the release carries no benchmark comparison, inference cost or accuracy figure, so there is no way to tell whether its 20-second rollouts beat existing simulators.

The Institute of Foundation Models at MBZUAI released PAN 2, a world model — one trained to predict what happens next in an environment — under Apache 2.0, so it can be used and modified freely. It pairs an LLM planner with a 29B-parameter diffusion transformer, the video-generation kind of network that refines noise into frames, and renders up to 20 seconds of a predicted action outcome as video. The announcement gave no evaluation results, training data, compute cost or inference speed.

Key facts

  • ·PAN 2 is released under the Apache 2.0 licence by the Institute of Foundation Models at MBZUAI. source
  • ·The model couples an LLM planner with a 29B-parameter diffusion transformer. source
  • ·It simulates up to 20 seconds of action outcomes as video. source
  • ·It is presented as a world model for physical AI. source

What the sources say

  • MarkTechPost — Short release write-up naming the architecture, parameter count, licence and rollout length, with no measurements.

Sources

The original reporting. Follow these — they did the work.

← the wire