The Institute of Foundation Models (IFM) at MBZUAI has released PAN 2, an open-source 29B physical AI world model under the Apache 2.0 license. Announced on November 10, 2026, the model allows robotics and software agents to simulate how physical scenes change based on text or image inputs, enabling simulative reasoning prior to real-world execution. It succeeds the original PAN model released in November 2025.

Built on IFM’s Generative Latent Prediction architecture, PAN 2 splits simulation into conceptual and detailed spatial reasoning. IFM reported that removing the model’s underlying LLM caused full-plan task completion to fall from 34.0% to 0.0%, demonstrating the role of textual reasoning in physical modeling. IFM also introduced a conceptual reasoner skill for Claude Code and Codex, alongside an optional guardrail that reduced unsafe text-to-video outputs from 33.9% to 1.5% in tests.

The 29B model requires loading the full parameters in BF16 per GPU worker and supports execution via Hugging Face Diffusers, vLLM, FastVideo, or native scripts. Open access to model weights, training code, and inference code provides developers with infrastructure to test and validate agentic plans in simulated environments.

Why it matters

  • Open-sourcing a 29B physical world model under Apache 2.0 allows robotics startups to test planning in simulation without building proprietary physical engines.

  • Ablation data proves LLM conceptual reasoning is essential for video-based simulation, offering guidance on architecture design for spatial AI systems.

  • Deploying the full BF16 model requires dedicated GPU infrastructure, making optimization frameworks like vLLM critical for real-time robotic deployment.

Source: marktechpost.com