MBZUAI's IFM open-sources PAN 2, a 29B world model that simulates action outcomes
LinusEkenstam · x · 2026-11-13
MBZUAI's Institute of Foundation Models unveiled PAN 2, a general-purpose world model: given a scene (optional image + text) and text-described actions, it simulates up to 20 seconds of video showing how the world changes — enabling "what if?" reasoning for agents.
Key points:
- Jenga demo: pulling a middle block topples the tower; slowly pulling a top block leaves it standing
- Agents can simulate candidate actions, pick the best outcome, then act in the real world; in IFM's tests an agent discovered a hidden property of the world purely through simulated interaction
- Architecture: Generative Latent Prediction (GLP) — an LLM handles conceptual reasoning and emits structured, time-grounded event descriptions; a 29B-parameter MMDiT diffusion backbone trained from scratch on hundreds of millions of annotated videos predicts perceptual detail; Wan 2.2 VAE decodes the latents
Weights, training and inference code are all open under Apache 2.0 on Hugging Face.
Related event: MBZUAI Open-Sources PAN 2, a World Model That Simulates Action Outcomes(2 posts)→
More from Multimodal
- Dev one-shots ASMR-ifying any video using ElevenLabs sound effects — austin_malerba · 2026-11-13
- PlaygroundArt.AI festival clip wows Reddit with lifelike motion and hair detail — shahryar2025 · 2026-11-13
- Full Restaurant Brand Visuals for ~24 Higgsfield Credits: A Step-by-Step Workflow — PrajwalTomar_ · 2026-11-13
- Structured SillyTavern-to-ComfyUI prompts: interaction beats pose for Krea 2 and H3 — Meinungsbild · 2026-11-13
- Gemini 4 Hydrogen builds 400K-voxel Chinese mountain temple in Blender, zero manual input — iamfakhrealam · 2026-11-13
- Open-source browser-based bbox layout builder for MOSS-Image, fully offline with JSON export — Rex-Racer9 · 2026-11-13