FULL STORY

PAN 2: MBZUAI's Open-Source World Model from Launch to Hands-On

MBZUAI's IFM released the open-weight world model PAN 2 on Nov 11. Community demos highlighted its physics consistency, and later hands-on tests compared its scene fidelity to GTA-level graphics.

2026-11-11 ~ 2026-11-22 · 4 episodes · 29 posts

Episode 1 · MBZUAI IFM Open-Sources PAN 2, a Foundation World Model for Physical Intelligence (2026-11-11, 16 posts)

On November 11, MBZUAI's Institute of Foundation Models (IFM, Eric Xing's team) released PAN 2, an open-weight foundation world model for physical intelligence. Instead of predicting the next token like a language model, PAN 2 predicts the next world—how the world evolves after each action, grounded in time, physics, and space. Its core capability is simulation-based reasoning: agents can rehearse what would happen in imagination, simulating multiple candidate actions before deciding.

Confirmed

  • Architecture is based on generative latent prediction (GLP) with a two-level division of labor: a conceptual level where an LLM reasons about what happens next and how to achieve goals, and a generative level with a 29B-parameter latent diffusion model (dual-stream multimodal diffusion transformer, excluding the inference backbone) that produces concrete world states.
  • Usage: given a current observation (optionally image + text) and a text-described action, it outputs up to 20 seconds of 704x1248@24fps (720p) video simulating the scene's evolution under time-anchored action plans and camera motion; text-to-video and image-to-video are supported.
  • Per @rohanpaulai, evaluations report 75.1% single-step accuracy and 34% multi-step planning task completion; demos include the 29B model simulating robot laundry-folding and other physical tasks.
  • Trained on millions of hours of egocentric and robot data.
  • The team lists six capability dimensions for world models: obeying physical laws, historical coherence, persistence out of sight, constraint satisfaction, viewpoint invariance, and more.
  • Highly open: Apache 2.0 license, weights on Hugging Face, integrated into Hugging Face Diffusers; per IFM's Reddit release post (relayed by @Aggressive-Zebra-949), training code is also released; the vLLM team announced day-zero support. Paper: arXiv:2507.05169.

Why it matters

  • PAN 2 moves world models from concept to a usable open-source foundation for robotics and physical intelligence research, offering a general substrate for pre-action simulation. Apache 2.0 licensing, Diffusers integration, and day-zero vLLM support keep deployment and reproduction barriers low, enabling rapid community adoption.

Episode 2 · PAN 2 World Model Demonstrates Physics, Temporal and Spatial Consistency (2026-11-11, 3 posts)

IFM's newly released world model PAN 2 was demoed handling glass refraction, sequential causal ordering of events, and spatial consistency across camera motion.

Episode 3 · MBZUAI Open-Sources PAN 2, a World Model That Simulates Action Outcomes (2026-11-13, 2 posts)

MBZUAI's IFM released PAN 2, a general world model that generates up to 20 seconds of video simulating how the world would change given an action, with weights and inference code open-sourced on Hugging Face under Apache 2.0.

Episode 4 · Open-source world model PAN 2 stress-tested with GTA-style chaos physics (2026-11-22, 8 posts)

On November 22, while waiting for GTA 6 to download, Linus Ekenstam ran a series of intensive hands-on tests on the open-source world model PAN 2, sharing results across multiple posts. He said the model's physics felt so real it was like driving in Miami, with an overall quality comparable to GTA-grade driving scenes.

Confirmed

  • Chaotic physics scenarios: demonstrated "what if" moments such as driving into the sea, a police car chasing a motorcycle through slippery alleys and smashing cardboard boxes (the motorcycle stayed consistent after the collision), a monster truck crushing a row of cars, and a motorcycle jumping over a drawbridge as it was opening.
  • Temporal evolution and consistency: a dockside night scene slowly transitioned to dawn with a boat emerging from the water, showing long-horizon consistency; as the camera followed a jet ski across the bay, the city skyline stayed consistent throughout.
  • Real-time interactive world generation: demonstrated a generative "game world" in real time, where the player skips a mission, jumps the drawbridge on a motorcycle, and crosses the bay on a jet ski—the model maintained scene coherence through continuous interaction and camera movement, rather than pre-rendered fixed content.
  • Fun finding: the model can also generate scenes you'd never see in GTA—a driver dutifully waiting at a red light.

Why it matters

  • Playing off Rockstar's meme that "GTA 6 is the first GTA that encourages empathy," Ekenstam ran chaos tests that vividly showcase a world model's capability boundaries under collision physics, long-horizon consistency, and large-scale camera motion; he also noted that chaotic scenes (like a police chase through slippery alleys hitting cardboard boxes) remain a hard case for world models. Demos like this hint at a potential future of generative real-time game worlds.