Open world model PAN 2 tops multi-step plan completion at 34%, simulates action outcomes
rohanpaul_ai · x · 2026-11-12
IFM released PAN 2, an open foundation world model for physical intelligence: describe an action and it generates up to 20 seconds of video predicting what physically happens next, letting agents simulate "what if?" before acting. Architecture splits the job in two — a language model works out the plot, then a 29B video model renders it.
Benchmarks (own 6D-PI tests plus Physics-IQ and RBench):
- 75.1% single-step completion; 34.0% full multi-step plan completion, the highest among tested models
- Beats open rivals on plan completion: MiniMax-H3 32.5%, Cosmos 3 Super 9.5%, LTX-2.5 4.0%; closed models FLUX 3 at 26.8% and Seedance 2.5 at 24.5%
- Physics-IQ Verified: 49.38, first among open models and third overall
Commentator framing: agents learn by trial and error and the error is the expensive part — world models move that error into simulation.
More from Multimodal
- Reddit user slams MiniMax for confusing model naming that pollutes search results — stook · 2026-11-12
- Browsable library tests MiniMax H3 Image across 288 style prompts with one base photo — Numerous_Access_8008 · 2026-11-12
- Midjourney --sref 2782468810 turns McDonald's into ukiyo-e prints — sergeantsref · 2026-11-12
- Google Carbon demo: simulating a stained-glass Jenga tower's light and shatter — AI_Andrew · 2026-11-12
- Claude Opus 5.5 composes game soundtracks for three distinct level vibes — wollymonjac · 2026-11-12
- Hero's Journey: open-source six-step AI short-film workflow, free to fork — neuroform · 2026-11-12