MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI

World models let a robot or agent rehearse an action in simulated video before trying it in the real world. MBZUAI's Institute of Foundation Models released PAN 2, a 29-billion-parameter model that takes a scene and a text action and renders what happens next, with weights, training code and inference code under Apache 2.0.

Fully open matters here. Competing models from MiniMax, NVIDIA and Lightricks ship under community or custom licenses, and several don't disclose training code. PAN 2 gives researchers something they can actually inspect, retrain and build on.

We like the design choice at the center of this. PAN 2 splits simulation in two: a language model writes out the next world state as structured events and actions, and the 29B diffusion model renders it as video. IFM's own ablation shows how much the planner carries the load. Remove it, and full-plan completion drops from 34% to zero. That's a useful finding for the whole field: reasoning about what should happen and rendering what it looks like are separate jobs.

Now the sober part. All benchmarks here are self-reported by IFM. PAN 2 leads six tested models on completing full multi-step plans at 34%, but MiniMax-H3 beats it on single actions, overall capability scores and human preference. And 34% means most long plans still miss a step. Hardware is heavy too: each GPU worker loads the full model, and IFM hasn't published memory needs or generation times.

I'd also note the guardrail numbers, because IFM was candid about them. Its recommended safety filter cut unsafe outputs on harmful video requests from 33.9% to 1.5%, but also blocked 11.5% of benign ones. Anyone deploying this should run the guardrail and expect some false refusals. Open weights are a gift to research, and they also mean nobody can force the guardrail on.

Our read: this is a real contribution to open physical AI and a fair benchmark of how far there is to go. Simulating before acting is the right instinct for robots. A simulator that's wrong two-thirds of the time on long tasks is a research tool, not a safety case.

PAN 2 succeeds PAN, released in November 2025, and follows the Generative Latent Prediction architecture from IFM's Critique of World Models paper. It outputs 704×1248 video at 24 frames per second for up to 20 seconds and has no audio.