MBZUAI releases PAN 2, a general-purpose world model for simulative agent reasoning
kimmonismus · x · 2026-11-11
MBZUAI's Institute of Foundation Models has open-sourced PAN 2, a general-purpose world model (Apache 2.0, arXiv:2507.05169).
- Capability: given an observation (optional image + text) and actions described in text, it simulates the outcome as a video up to 20 seconds, enabling "simulative reasoning" — an agent can imagine candidate actions' outcomes before acting.
- Architecture: follows the Generative Latent Prediction (GLP) design from Critique of World Models. An LLM handles conceptual reasoning, writing the next world state as a structured, time-grounded description (events, actor-bound actions, camera motion); a from-scratch 29B-parameter MMDiT diffusion backbone predicts video latents; Wan 2.2's VAE decoder renders them.
- Training: hundreds of millions of videos annotated in a unified format.
- Demo highlight: consistent scenes across moving viewpoints — PAN 2 tracks a paper plane thrown off a balcony all the way down while plants, sofas and floor stay unchanged. Commentator kimmonismus notes that if agents can test plans in simulation first, much slow, costly real-world trial and error moves out of reality.
More from Models
- Timed eval task asks users to deliberately make Claude fail at real work in 12 minutes — Icy-Measurement10 · 2026-11-22
- Tesla FSD 15's real-time view isn't 3D rendering — it's the AI model's own synthetic perception — mark_k · 2026-11-22
- Gemini Black Friday half-price: JP-region trick gets Pro annual plan for $91.54 — huangyun_122 · 2026-11-22
- Microsoft's USB4 RDMA lets mini PCs pool memory to run Qwen3-235B — JFPuget · 2026-11-22
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22