MBZUAI’s IFM releases open 29B PAN 2 world model
Original title: MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI
Apache 2.0 weights and code ship; IFM reports 34% full-plan completion, ahead of MiniMax-H3 on multi-step tasks.
The Institute of Foundation Models at MBZUAI released PAN 2 on November 10, 2026. The 29B diffusion transformer takes a scene plus a text action and generates video up to 20 seconds. Weights, training code and inference code are under Apache 2.0. Benchmark figures are IFM’s own; MiniMax-H3 still leads several single-action and preference scores.
Unconfirmed: Benchmark scores are self-reported by IFM; the blog cites different scores for some rivals.
- Output is 704×1248 video at 24 fps, up to 20 seconds.
- Default planner uses Kimi K3; the video model is a 29B MMDiT.
- Without conceptual reasoning, full-plan completion falls from 34.0% to 0.0%.
- In IFM tests, pan2-guardrail cut unsafe text-to-video outputs from 33.9% to 1.5%.
- Organizations
- MBZUAIInstitute of Foundation Models
- AI models
- PAN 2Kimi K3Qwen3.5-9BMiniMax-H3
From the original article
8 min read at marktechpost.com
The Institute of Foundation Models at MBZUAI released PAN 2, an Apache 2. 0 world model that couples an LLM planner with a 29B diffusion transformer to simulate up to 20 seconds of action outcomes as video. The post MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI appeared first on MarkTechPost.
By Michal Sutter · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.