PAN 2

PAN 2 is an open foundation world model for physical intelligence, announced on 10 November 2026 by the Institute of Foundation Models (IFM) at MBZUAI. On the official product page and in the announcement from @IFM_AI, the lab's line is: a language model predicts the next word; PAN 2 predicts the next world — what happens after each action, grounded in time, physics and space — so people and AI agents can simulate “what if?” before they act. MBZUAI's account @mbzuai posted congratulations and said the release includes model weights plus training and inference code.

The official page describes the method as generative latent prediction (GLP), with two reasoning backbones. A multimodal LLM does conceptual reasoning: it works out what happens and when, including the timing of actions and camera motion. A 29-billion-parameter latent diffusion model, trained from scratch, does perceptual reasoning: it fills in geometry, appearance, and how things move. The page illustrates the design with examples such as a surfer's pop-up timing, milk versus ink versus oil poured into water, and a camera that can orbit a scene without rewriting the world it shows.

IFM presents PAN 2 as a rehearsal space for agents and humans before real-world acting, not merely a video generator optimized for looks. A Hugging Face repo is listed at IFM/PAN2. License, hardware needs, and exact checkpoints should be taken from the official page and the repository card.

Highlights:

• Official page: https://ifm.ai/pan/ ;
• Official announcement (10 November 2026): https://x.com/IFM_AI/status/2120199814696828942 ;
• MBZUAI congratulations: https://x.com/mbzuai/status/2120202298119045374 ;
• Hugging Face: https://huggingface.co/IFM/PAN2 ;
• This is a new generation name on IFM's world-model line. It is not the same thing as unrelated “PAN” shared-task or project names.

Why it matters

If a world model is going to serve as a rehearsal engine for agents, visual realism is not enough: action timing, physical consequences, and camera motion must not scramble the underlying world. IFM is publishing a named split between conceptual planning (who does what, and when) and perceptual rendering (how it looks and moves), together with weights and training / inference code, which gives a concrete place to reproduce and fine-tune.

The demos and capability claims come from IFM / MBZUAI. We have not independently verified physical consistency or any comparison against other world models. Treat this as an open research release you can download and inspect, not as a guarantee that it can replace real physics simulation in production.

Limits

Official material is generated-video and interactive examples, not a claim that arbitrary real-robot closed-loop control is ready. The Hugging Face repository may predate this public announcement; use the 10 November product page and official post as the public launch, and check the current repo docs for APIs, license, and dependencies. Do not confuse PAN 2 with other projects or evaluation tasks that also use the name “PAN.”

1. Read the official page https://ifm.ai/pan/ and the announcement https://x.com/IFM_AI/status/2120199814696828942 ;
2. See the MBZUAI congratulations post https://x.com/mbzuai/status/2120202298119045374 for the statement that weights and training / inference code are released together ;
3. Check https://huggingface.co/IFM/PAN2 for the current checkpoint and usage notes; follow the repository for license and runtime requirements ;
4. Treat PAN 2 as a starting point for open world-model research, not as a production-grade physics simulator.