PAN 2 Opens World-Model Training

PAN 2 matters less as a polished video generator than as an open sandbox for agents to rehearse decisions before acting.

• –Separating conceptual planning from perceptual rendering makes simulations more inspectable and extensible
• –Agents can compare counterfactual futures instead of executing the first plausible action
• –IFM reports 75.1% step accuracy but only 34% full-plan completion, underscoring major reliability gaps
• –The 29B model and multi-stage pipeline create a substantial hardware barrier for experimentation
• –Apache 2.0 licensing could accelerate research into embodied agents and learned simulators