MBZUAI releases open PAN 2 world model; vLLM adds day-one video generation support
Article timeline
The articles in this story. Times are ET.
SantiagoXSantiago explains how PAN 2 lets agents simulate actions before choosingAISantiago explains that PAN 2, a world model from the Institute of Foundation Models, lets an agent simulate each possible action and pick the best outcome. He describes its Generative Latent Prediction design, with an LLM handling conceptual reasoning and a 29B LDM rendering the predicted scene. He says the weights, training code, and inference code are released under Apache 2.0.
Rohan PaulXMBZUAI releases PAN 2, an open world model with simulative reasoning for physical intelligenceAIMBZUAI's Institute of Foundation Models released PAN 2, an open foundation world model for physical intelligence, built for simulative reasoning. It pairs an LLM that plans actions and camera motion as text with a 29B diffusion model trained from scratch that renders the video. With both parts, PAN 2 completes 75.1% of steps correctly and finishes 34.0% of plans in full, the highest among the models IFM tested, versus 8.7% of steps without the LLM.
vLLMOfficialvLLM-Omni adds day-one support for PAN 2 video generationAIvLLM-Omni adds Day-0 support for PAN 2 text-to-video and image-to-video generation, making the open-weight world model available through an open inference stack. The post describes PAN 2 as an open-weight world model built to predict how scenes evolve as actions unfold, grounded in time, physics and space, and credits the PAN 2 team for the work.