The Institute of Foundation Models (IFM) at MBZUAI released PAN 2, an open world model for physical intelligence. The PAN 2 world model takes a scene and a text action and simulates the outcome as video. IFM announced it on X on November 10, 2026. Weights, training code and inference code ship under Apache 2.0. For robotics and agent developers, the pitch is simple: test a plan in simulation before acting on it.
TL;DR
- Size: 29B-parameter diffusion transformer (MMDiT), dense. Inputs: text plus an optional first frame. Output: 704×1248 video at 24 fps, up to 20 s.
- Runs on: Data-center GPUs. Each worker holds the full 29B model in BF16, with block offloading as fallback.
- Performance: Leads 6 tested models on full plans. Trails MiniMax-H3 on single actions and human preference.
- Best: 34.0% full-plan completion on 100 multi-step tasks, the highest of all models tested (self-reported).
- Worst: About 2 in 3 plans miss a step. Physics-IQ system default: 37.48 vs 42.91 for MiniMax-H3.
- Bottom line: Best: a fully open world model that leads on multi-step plans. Worst: long plans still fail often, and hardware needs are high.
What is PAN 2?
PAN 2 is a general-purpose world model from IFM that simulates how a scene changes after an action. It follows the Generative Latent Prediction (GLP) architecture from IFM’s Critique of World Models paper. It succeeds PAN, released in November 2025.
The goal is simulative reasoning. An agent proposes plans, PAN 2 simulates each one, and the agent acts on the best.
How does PAN 2 work?
PAN 2 splits simulation into 2 kinds of reasoning:
- Conceptual reasoning: An LLM reads the observation and the requested actions. It writes the next world state as structured JSON: events, timed actions bound to actors, and camera motion. Any instruction-following LLM can fill this role. IFM’s report uses Kimi K3 by default.
- Perceptual reasoning: A 29B MMDiT predicts the video latents. IFM trained it from scratch on hundreds of millions of annotated videos. A frozen Qwen3.5-9B text encoder conditions it. The Wan 2.2 VAE decoder renders the final frames.
IFM’s ablation shows the LLM matters. With no conceptual reasoning, full-plan completion drops from 34.0% to 0.0%. Step completion falls from 75.1% to 8.7%. A format-only rewrite recovers part of the gap, reaching 27.8% full-plan completion.
IFM also dropped gameplay, animation and simulation footage after early training. A 9B test model trained without it won 58% of blinded votes.
‘+v+’
‘;box.appendChild(row);
var mn=m.min||0,w=Math.max(2,(r[1]-mn)/(m.max-mn)*100);
requestAnimationFrame(function(){requestAnimationFrame(function(){row.querySelector(‘.fill’).style.width=w+’%’})});});
document.getElementById(‘mtpNote’).textContent=m.note+(m.min?’ Axis starts at ‘+m.min+’.’:”);setTimeout(post,60);}
// ablation
var A=[[45.1,8.7,0.0],[78.2,70.7,27.8],[84.6,75.1,34.0]],abl=2;
function setAbl(i){abl=i;R.querySelectorAll(‘#mtpAblSeg .chip’).forEach(function(c){c.classList.toggle(‘on’,+c.dataset.a===i)});
[0,1,2].forEach(function(j){var el=document.getElementById(‘ab’+j),to=A[i][j],from=parseFloat(el.textContent)||0,t0=null;
function tick(ts){if(!t0)t0=ts;var p=Math.min(1,(ts-t0)/600),v=from+(to-from)*p;el.innerHTML=v.toFixed(1)+’%‘;if(p<1)requestAnimationFrame(tick)}
requestAnimationFrame(tick);document.getElementById(‘ib’+j).style.width=to+’%’;});}
R.querySelectorAll(‘#mtpAblSeg .chip’).forEach(function(c){c.onclick=function(){setAbl(+c.dataset.a)}});
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
