MBZUAI's IFM releases PAN 2, an open 29B world model
Start here
MBZUAI's Institute of Foundation Models (IFM) released PAN 2, a 29B-parameter open world model that takes a scene plus a text action and simulates the outcome as video. Weights, training code and inference code ship under Apache 2.0; it runs on data-center GPUs, with each worker holding the full model in BF16 and block offloading as a fallback, and IFM strongly recommends a separate pan2-guardrail. The system pairs an LLM planner that writes structured JSON world states with a 29B MMDiT diffusion transformer conditioned by a frozen Qwen3.5-9B encoder. IFM reports 34.0% full-plan completion on 100 multi-step tasks, ahead of MiniMax-H3 at 32.50%, while trailing MiniMax-H3 on single actions and human preference (Elo 1595 vs 1620). An IFM ablation shows that removing the LLM planner drops full-plan completion to 0.0% and step completion from 75.1% to 8.7%.
Generated from reports · updated 14 h ago
Report timeline
Follow the reports to see the event from different sides.
- MarkTechPostMBZUAI's IFM Releases PAN 2, an Open 29B World Model for Physical AI
MBZUAI's Institute of Foundation Models released PAN 2, a 29B-parameter open world model that takes a scene plus a text action and simulates the outcome as video, with weights, training code and inference code under Apache 2.0. It pairs an LLM planner that writes structured JSON world states with a 29B MMDiT diffusion transformer conditioned by a frozen Qwen3.5-9B encoder, and IFM reports 34.0% full-plan completion on 100 multi-step tasks, ahead of MiniMax-H3 at 32.50%, while trailing MiniMax-H3 on single actions and human preference (Elo 1595 vs 1620). IFM's ablation shows removing the LLM planner drops full-plan completion to 0.0% and step completion from 75.1% to 8.7%.
Heat for this event
Not enough continuous observations yet to draw a trend.