RealityHackerOpen in RealityHacker ⇢
Models · Multimodal Models · published 2026-11-11T00:00:00+00:00 · via MarkTechPost

MBZUAI's IFM Unveils PAN 2: A 29B-Parameter Open World Model for Simulating Physical Actions

PAN 2, developed by MBZUAI's Institute of Foundation Models, is an open-source world model with a 29B-parameter diffusion transformer that integrates an LLM planner to simulate up to 20 seconds of physical action outcomes in video form. The model operates under the Apache 2.0 license, enabling broad access for research and development in physical AI. It represents a significant advancement in multimodal simulation of real-world interactions using large-scale foundation models.

Expanded Detail

MBZUAI’s Institute of Foundation Models has introduced PAN 2, a 29-billion-parameter open-world model designed to simulate physical actions over up to 20 seconds in video form. Built on a diffusion transformer architecture and enhanced with an integrated large language model (LLM) planner, PAN 2 represents a leap in multimodal simulation by combining reasoning and visual prediction at scale. The model is released under the permissive Apache 2.0 license, making it freely available for research and development. This open-source approach enables widespread access to advanced physical AI simulation tools, potentially accelerating innovation across robotics, autonomous systems, and digital twin technologies.

The development of PAN 2 builds on growing interest in foundation models capable of understanding and predicting real-world dynamics. Prior work in multimodal AI has focused on static image or text understanding, but recent advances have emphasized dynamic simulation—especially in environments where physical laws govern outcomes. MBZUAI, a research institution based in the UAE dedicated to foundational AI, has positioned itself at the forefront of this shift by investing in large-scale models that bridge language, vision, and action prediction. The organization’s goal appears to be fostering global collaboration through open access while establishing regional leadership in AI innovation.

What remains uncertain is how PAN 2 will perform across diverse physical scenarios beyond controlled test environments, and whether its 20-second simulation window can scale effectively to longer or more complex sequences. Additionally, the model’s reliance on an LLM planner raises questions about computational efficiency and real-time deployment. Future developments may include integration with robotics platforms, expansion into multi-agent simulations, or adaptation for training in virtual environments.

Context

PAN 2 could significantly influence how researchers and engineers prototype physical systems, particularly in robotics and autonomous vehicle development, by reducing reliance on costly physical testing. Its open-source nature may democratize access to high-fidelity simulation tools, enabling startups and academic institutions with limited resources to participate in cutting-edge AI research. This shift may accelerate innovation cycles across industries that depend on predictive modeling—such as manufacturing, logistics, and urban planning. The broader impact may ripple into education and policy, as accessible simulation tools could reshape how physical systems are taught and regulated. Comparable to the release of open-source models like Llama or Stable Diffusion, PAN 2 may catalyze a new wave of community-driven development in physical AI. However, its real-world deployment remains uncertain—questions about safety, bias, and unintended consequences in simulated environments could emerge as adoption grows. What to watch for next is how developers adapt PAN 2 beyond research settings, and whether it becomes a standard benchmark in multimodal simulation.

Expanded detail and Context are AI-generated analysis; the linked article remains the authoritative source.
Read the full article at MarkTechPost →
Related stories
DeepMind Unveils AlphaGenome 2 to Predict Cell-Type-Specific Genetic Variant Impacts · Multimodal Models
Open-Source Text-to-Speech Model Delivers Real-Time Voice Cloning with Efficient Parameter Use · Large Language Models
This summary is Al-enhanced to contain extended analysis and broader social context. The original is {NAME); the linked article is the authoritative source. Original headline: “MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI.” Browse more stories.