Sakhanda Wire
NVDA $154.51 -1.04% MSFT $506.49 -2.22% GOOGL $281.27 -0.86% META $621.51 -0.78% AMZN $256.90 -0.33%
← Back to the news

MBZUAI releases open PAN 2 world model for physical AI

The Institute of Foundation Models (IFM) at MBZUAI has released PAN 2, an open world model designed to simulate how a scene changes after a specified action. It accepts text and, optionally, an initial image, then generates video showing the predicted result. The model, weights, training code and inference code are available under the Apache 2.0 license.

PAN 2 combines an instruction-following large language model with a 29B-parameter MMDiT diffusion transformer. The language model converts an observation and requested action into a structured description of events, timed actions and camera movement. The diffusion model then generates video at 704×1248 resolution and 24 frames per second, for up to 20 seconds. IFM’s default setup uses Kimi K3 for conceptual reasoning, Qwen3.5-9B as a text encoder and the Wan 2.2 VAE decoder.

In IFM’s self-reported tests, PAN 2 completed every step in 34.00% of 100 multi-step tasks, compared with 32.50% for MiniMax-H3. Individual step completion was 75.10%, close to MiniMax-H3’s 75.12%. The system requires data-center GPUs, with each worker holding the full model in BF16. Without the language-model reasoning stage, full-plan completion fell to 0.0% in an ablation.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news