The Institute of Foundation Models at MBZUAI has released PAN 2, an open world model that takes a scene and a text action and simulates the result as video, MarkTechPost reports. Weights, training code and inference code are Apache 2.0. The pitch for robotics and agents: try a plan in simulation before you act on it.
- A 29B-parameter diffusion transformer. Optional first frame plus text in, up to 20 seconds of 704×1248 video at 24 fps out.
- It needs data-center GPUs; each worker holds the full 29B model in BF16.
- On IFM’s own tests it leads the models they compared on full-plan completion (34%), and trails MiniMax-H3 on some single-action and preference scores. About two in three long plans still miss a step.
My take
A fully open world model is useful even if the plan scores aren’t perfect. The hardware bar is high, and the numbers are self-reported, so read the caveats.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
God Bless, and May the Source Be with You!
