MBZUAI’s IFM Releases PAN 2, an Open 29B World Mannequin for Bodily AI

The Institute of Basis Fashions (IFM) at MBZUAI launched PAN 2, an open world mannequin for bodily intelligence. The PAN 2 world mannequin takes a scene and a textual content motion and simulates the result as video. IFM introduced it on X on November 10, 2026. Weights, coaching code and inference code ship below Apache 2.0. For robotics and agent builders, the pitch is straightforward: take a look at a plan in simulation earlier than appearing on it.

TL;DR

Dimension: 29B-parameter diffusion transformer (MMDiT), dense. Inputs: textual content plus an non-compulsory first body. Output: 704×1248 video at 24 fps, as much as 20 s.
Runs on: Information-center GPUs. Every employee holds the total 29B mannequin in BF16, with block offloading as fallback.
Efficiency: Leads 6 examined fashions on full plans. Trails MiniMax-H3 on single actions and human desire.
Finest: 34.0% full-plan completion on 100 multi-step duties, the very best of all fashions examined (self-reported).
Worst: About 2 in 3 plans miss a step. Physics-IQ system default: 37.48 vs 42.91 for MiniMax-H3.
Backside line: Finest: a completely open world mannequin that leads on multi-step plans. Worst: lengthy plans nonetheless fail usually, and {hardware} wants are excessive.

What’s PAN 2?

PAN 2 is a general-purpose world mannequin from IFM that simulates how a scene modifications after an motion. It follows the Generative Latent Prediction (GLP) structure from IFM’s Critique of World Fashions paper. It succeeds PAN, launched in November 2025.

The objective is simulative reasoning. An agent proposes plans, PAN 2 simulates each, and the agent acts on the most effective.

How does PAN 2 work?

PAN 2 splits simulation into 2 sorts of reasoning:

Conceptual reasoning: An LLM reads the remark and the requested actions. It writes the following world state as structured JSON: occasions, timed actions certain to actors, and digicam movement. Any instruction-following LLM can fill this function. IFM’s report makes use of Kimi K3 by default.
Perceptual reasoning: A 29B MMDiT predicts the video latents. IFM skilled it from scratch on lots of of tens of millions of annotated movies. A frozen Qwen3.5-9B textual content encoder situations it. The Wan 2.2 VAE decoder renders the ultimate frames.

IFM’s ablation exhibits the LLM issues. With no conceptual reasoning, full-plan completion drops from 34.0% to 0.0%. Step completion falls from 75.1% to eight.7%. A format-only rewrite recovers a part of the hole, reaching 27.8% full-plan completion.

IFM additionally dropped gameplay, animation and simulation footage after early coaching. A 9B take a look at mannequin skilled with out it received 58% of blinded votes.

<div class="pane on" id="p-how">
<div class="pipe">
<div class="node" data-i="0"><div class="n">STEP 1</div><div class="t">Observe</div><div class="s">Elective picture + textual content motion</div></div>
<div class="node" data-i="1"><div class="n">STEP 2</div><div class="t">Conceptual</div><div class="s">LLM writes the following state</div></div>
<div class="node" data-i="2"><div class="n">STEP 3</div><div class="t">Encode</div><div class="s">Frozen Qwen3.5-9B</div></div>
<div class="node" data-i="3"><div class="n">STEP 4</div><div class="t">Perceptual</div><div class="s">29B MMDiT predicts latents</div></div>
<div class="node" data-i="4"><div class="n">STEP 5</div><div class="t">Render</div><div class="s">Wan 2.2 VAE decodes video</div></div>
</div>
<div class="stream"><i id="mtpFlow"></i></div>
<div class="element" id="mtpDetail"></div>
<div class="ctrl">
<button class="btn" id="mtpPrev"> Prev</button>
<button class="btn pri" id="mtpNext">Subsequent </button>
<button class="btn" id="mtpPlay"> Auto-play</button>
<span class="trace">Click on any step to leap</span>
</div>
</div>

<div class="pane" id="p-abl">
<div class="seg" id="mtpAblSeg">
<button class="chip" data-a="0">None: motion despatched straight to diffusion</button>
<button class="chip" data-a="1">Format-only rewrite</button>
<button class="chip on" data-a="2">Full conceptual reasoning</button>
</div>
<div class="abl">
<div class="card"><div class="okay">6 capabilities</div><div class="v" id="ab0">84.6<small>%</small></div><div class="mini"><i id="ib0"></i></div></div>
<div class="card"><div class="okay">Step completion</div><div class="v" id="ab1">75.1<small>%</small></div><div class="mini"><i id="ib1"></i></div></div>
<div class="card"><div class="okay">Full-plan completion</div><div class="v" id="ab2">34.0<small>%</small></div><div class="mini"><i id="ib2"></i></div></div>
</div>
<p class="observe">IFM ablation with Kimi K3 because the reasoner; the diffusion mannequin stays mounted. With out the LLM, no multi-step plan completes. Supply: IFM technical weblog.</p>
</div>

<div class="pane" id="p-run">
<div class="grid">
<div class="spec"><div class="okay">Generator</div><div class="v">29B MMDiT, BF16, skilled from scratch</div></div>
<div class="spec"><div class="okay">Duties</div><div class="v">Textual content-to-video, image-to-video</div></div>
<div class="spec"><div class="okay">Output</div><div class="v">704×1248 at 24 fps, as much as 20 s (481 frames)</div></div>
<div class="spec"><div class="okay">Default clip</div><div class="v">241 frames, about 10 s</div></div>
<div class="spec"><div class="okay">Sampler</div><div class="v">50 steps · stream shift 9 · CFG 2.5</div></div>
<div class="spec"><div class="okay">Caption price range</div><div class="v">1,536 tokens onerous restrict</div></div>
<div class="spec"><div class="okay">Serving</div><div class="v">Diffusers · vLLM · FastVideo · generate.py</div></div>
<div class="spec"><div class="okay">GPU reminiscence</div><div class="v">Full 29B per employee; block offload fallback; VRAM not disclosed</div></div>
<div class="spec"><div class="okay">Security</div><div class="v">pan2-guardrail: unsafe T2V 33.9% → 1.5%</div></div>
<div class="spec"><div class="okay">License</div><div class="v">Apache 2.0: weights, coaching + inference code</div></div>
</div>
<pre>hf obtain IFM/PAN2 –include "transformer/*"
hf obtain Qwen/Qwen3.5-9B
hf obtain Wan-AI/Wan2.2-TI2V-5B-Diffusers</pre>
</div>

<div class="ft">
<span>Sources: <a href="https://huggingface.co/IFM/PAN2" goal="_blank" rel="noopener">HF mannequin card</a> · <a href="https://ifm.ai/weblog/pan2/" goal="_blank" rel="noopener">IFM weblog</a> · scores self-reported</span>
<span class="mtp">© Marktechpost</span>
</div>
</div>
<script>
(operate(){
var R=doc.getElementById(‘mtp-pan2′);
operate put up(){strive{dad or mum.postMessage({sort:’mtp-pan2-h’,h:R.offsetHeight+40},’*’)}catch(e){}}
// tabs
R.querySelectorAll(‘.tab’).forEach(operate(b){b.addEventListener(‘click on’,operate(){
R.querySelectorAll(‘.tab’).forEach(operate(x){x.classList.take away(‘on’)});
R.querySelectorAll(‘.pane’).forEach(operate(x){x.classList.take away(‘on’)});
b.classList.add(‘on’);doc.getElementById(‘p-‘+b.dataset.p).classList.add(‘on’);
if(b.dataset.p===’bench’)drawBars(cur);
if(b.dataset.p===’abl’)setAbl(abl);
setTimeout(put up,60);
})});
// pipeline
var D=[
[‘Observe: current world + action’,’Input is an optional starting image plus text, and actions described in text with a duration. Example: "Accelerate mangrove forest growth to protect the coast."’,”],
[‘Conceptual reasoning backbone (LLM)’,’An instruction-following LLM reasons over the scene and the action. It writes the next world state as structured, time-grounded JSON. IFM’s report uses Kimi K3 by default.’,'{ "events": [{"t":"0-3s","actor":"surfer","action":"paddles"},n {"t":"3s","actor":"surfer","action":"pops up"}],n "digicam": "sluggish pan proper" }’],
[‘Frozen text encoder’,’Qwen3.5-9B maps the structured caption to conditioning features. Captions have a soft limit of 1,480 and a hard limit of 1,536 tokens.’,”],
[‘Perceptual reasoning backbone (29B MMDiT)’,’A 29B diffusion transformer, trained from scratch on hundreds of millions of videos in the same caption format, fills in geometry, contact and motion. It predicts video latents conditioned on the state and the starting frame.’,”],
[‘Render the next observation’,’The Wan 2.2 VAE decoder turns latents into 704×1248 video at 24 fps, up to 20 s. An agent compares simulated outcomes and acts on the best plan.’,”]
];
var step=0,timer=null,nodes=R.querySelectorAll(‘.node’);
operate present(i){step=(i+5)%5;nodes.forEach(operate(n,okay){n.classList.toggle(‘on’,okay===step);n.classList.toggle(‘carried out’,okay<step)});
doc.getElementById(‘mtpFlow’).type.width=((step+1)/5*100)+’%’;
var d=D[step];doc.getElementById(‘mtpDetail’).innerHTML='<h4>’+d[0]+'</h4><p>’+d[1]+'</p>’+(d[2]?'<pre>’+d[2]+'</pre>’:”);setTimeout(put up,30);}
nodes.forEach(operate(n){n.addEventListener(‘click on’,operate(){cease();present(+n.dataset.i)})});
doc.getElementById(‘mtpNext’).onclick=operate(){cease();present(step+1)};
doc.getElementById(‘mtpPrev’).onclick=operate(){cease();present(step-1)};
var pb=doc.getElementById(‘mtpPlay’);
operate cease(){if(timer){clearInterval(timer);timer=null;pb.textContent=” Auto-play”}}
pb.onclick=operate(){if(timer){cease();return}pb.textContent=”❚❚ Pause”;timer=setInterval(operate(){present(step+1)},2600)};
present(0);
// bars
var M={
plan:{n:’Full-plan completion’,u:’%’,max:40,observe:’100 selfish duties, 456 steps. Share of simulations with each step carried out. † closed mannequin, single seed.’,d:[[‘PAN 2’,34.00,1],[‘MiniMax-H3’,32.50],[‘FLUX 3 †’,26.77,0,1],[‘Seedance 2.5 †’,24.49,0,1],[‘Cosmos 3 Super’,13.83],[‘LTX 2.5′,13.00]]},
step:{n:’Step completion’,u:’%’,max:80,observe:’Share of particular person plan steps carried out. PAN 2 and MiniMax-H3 are successfully tied.’,d:[[‘MiniMax-H3’,75.12],[‘PAN 2’,75.10,1],[‘FLUX 3 †’,72.44,0,1],[‘Seedance 2.5 †’,68.95,0,1],[‘Cosmos 3 Super’,57.23],[‘LTX 2.5′,55.77]]},
six:{n:’6 capabilities’,u:’%’,max:90,observe:’300 single-action circumstances (201 I2V, 99 T2V). MiniMax-H3 leads all 6 classes.’,d:[[‘MiniMax-H3’,87.20],[‘PAN 2’,84.48,1],[‘Seedance 2.5 †’,78.59,0,1],[‘FLUX 3 †’,77.41,0,1],[‘LTX 2.5’,72.25],[‘Cosmos 3 Super’,71.34]]},
piq:{n:’Physics-IQ Verified’,u:”,max:55,observe:’Leaderboard setting; real-take variance = 100. PAN 2 makes use of Claude Opus 5.5 as reasoner right here. LTX 2.5 not on leaderboard.’,d:[[‘FLUX 3 †’,51.11,0,1],[‘PAN 2’,49.38,1],[‘Cosmos 3 Super’,48.62],[‘Seedance 2.5 †’,42.43,0,1],[‘MiniMax-H3’,39.79],[‘Omni Flash 1.1 †’,35.34,0,1]]},
rb:{n:’RBench’,u:”,max:65,observe:’Robotics benchmark, unweighted imply x100.’,d:[[‘Omni Flash 1.1 †’,60.79,0,1],[‘FLUX 3 †’,57.96,0,1],[‘PAN 2’,57.04,1],[‘Seedance 2.5 †’,56.38,0,1],[‘MiniMax-H3’,56.16],[‘LTX 2.5’,51.55],[‘Cosmos 3 Super’,50.64]]},
elo:{n:’Human Elo’,u:”,max:1700,min:1250,observe:’Blind pairwise desire: 140 prompts, 13 reviewers, 1,400 comparisons. Open fashions solely.’,d:[[‘MiniMax-H3’,1620],[‘PAN 2 (GPT-6 Astra)’,1595,1],[‘PAN 2 (Claude Opus)’,1559,1],[‘Cosmos 3 Super’,1380],[‘LTX 2.5’,1346]]}
};
var cur=”plan”,seg=doc.getElementById(‘mtpSeg’);
Object.keys(M).forEach(operate(okay){var b=doc.createElement(‘button’);b.className=”chip”+(okay===cur?’ on’:”);b.textContent=M[k].n;b.onclick=operate(){cur=okay;seg.querySelectorAll(‘.chip’).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawBars(okay)};seg.appendChild(b)});
operate drawBars(okay){var m=M[k],field=doc.getElementById(‘mtpBars’);field.innerHTML=”;
m.d.forEach(operate(r){var row=doc.createElement(‘div’);row.className=”row”+(r[2]?’ me’:”);
var v=m.u===’%’?r[1].toFixed(2)+’%’:(okay===’elo’?String(r[1]):r[1].toFixed(2));
row.innerHTML='<div class="lab’+(r[3]?’ cl’:”)+’">’+r[0]+'</div><div class="trk"><div class="fill"></div></div><div class="val">’+v+'</div>’;field.appendChild(row);
var mn=m.min||0,w=Math.max(2,(r[1]-mn)/(m.max-mn)*100);
requestAnimationFrame(operate(){requestAnimationFrame(operate(){row.querySelector(‘.fill’).type.width=w+’%’})});});
doc.getElementById(‘mtpNote’).textContent=m.observe+(m.min?’ Axis begins at ‘+m.min+’.’:”);setTimeout(put up,60);}
// ablation
var A=[[45.1,8.7,0.0],[78.2,70.7,27.8],[84.6,75.1,34.0]],abl=2;
operate setAbl(i){abl=i;R.querySelectorAll(‘#mtpAblSeg .chip’).forEach(operate(c){c.classList.toggle(‘on’,+c.dataset.a===i)});
[0,1,2].forEach(operate(j){var el=doc.getElementById(‘ab’+j),to=A[i][j],from=parseFloat(el.textContent)||0,t0=null;
operate tick(ts){if(!t0)t0=ts;var p=Math.min(1,(ts-t0)/600),v=from+(to-from)*p;el.innerHTML=v.toFixed(1)+'<small>%</small>’;if(p<1)requestAnimationFrame(tick)}
requestAnimationFrame(tick);doc.getElementById(‘ib’+j).type.width=to+’%’;});}
R.querySelectorAll(‘#mtpAblSeg .chip’).forEach(operate(c){c.onclick=operate(){setAbl(+c.dataset.a)}});
window.addEventListener(‘load’,put up);window.addEventListener(‘resize’,put up);setTimeout(put up,300);
})();
</script>
</physique>
</html>
“>

(operate(){var f=doc.getElementById(‘mtp-pan2-frame’);window.addEventListener(‘message’,operate(e){if(e.information&&e.information.sort===’mtp-pan2-h’&&e.supply===f.contentWindow){f.type.top=e.information.h+’px’;}});})();

How does PAN 2 carry out on benchmarks?

All numbers beneath are self-reported by IFM. Its personal take a look at suite covers 6 capabilities throughout 300 single-action circumstances, plus 100 multi-step duties.

Full-plan completion: 34.00% vs 32.50% for MiniMax-H3, 26.77% for FLUX 3 and 24.49% for Seedance 2.5.
Step completion: 75.10%, stage with MiniMax-H3 at 75.12%.
6 capabilities total: 84.48% vs 87.20% for MiniMax-H3. MiniMax-H3 leads in all 6 classes.
Physics-IQ Verified (leaderboard setting): 49.38, first amongst open fashions. Closed FLUX 3 scores 51.11.
RBench: 57.04 vs 56.16 for MiniMax-H3 and 50.64 for Cosmos 3 Tremendous.
Human Elo: 1595 with a GPT-6 Astra reasoner, behind MiniMax-H3 at 1620.

PAN 2 vs MiniMax-H3, Cosmos 3 Tremendous and LTX-2.5

Function	PAN 2	MiniMax-H3	Cosmos 3 Tremendous	LTX-2.5
Developer	IFM (MBZUAI)	MiniMax	NVIDIA	Lightricks
Generator dimension	29B MMDiT	33B dense transformer (~13B in cacheable AdaLN branches)	64B Combination-of-Transformers	22B transformer (per checkpoint file names)
Planner / encoder	Any LLM + frozen Qwen3.5-9B encoder	Qwen3-VL-32B encoder; hosted Context-IR	Constructed-in autoregressive tower	Gemma-based textual content encoder
Max clip	20 s, 704×1248, 24 fps	15 s, 24 fps, 768p (2K through hosted module)	400 frames	Not disclosed
Audio	No	Sure, stereo	Sure	Sure
License	Apache 2.0	MiniMax H3 Group License	OpenMDW 1.1	LTX-2.x Group License
Coaching code	Sure	Not disclosed	Framework on GitHub	Not disclosed
Full-plan completion*	34.00%	32.50%	13.83%	13.00%
Physics-IQ Verified (leaderboard)*	49.38	39.79	48.62	Not reported
Human Elo*	1595	1620	1380	1346

*Scores from IFM’s PAN 2 mannequin card, run by IFM. The weblog reviews completely different Cosmos 3 Tremendous and LTX-2.5 plan scores.

How do you run PAN 2?

PAN 2 helps text-to-video and image-to-video. You possibly can run it by way of Hugging Face Diffusers, vLLM, FastVideo or the repo’s generate.py.

The default sampler makes use of 50 steps, stream shift 9 and steerage 2.5. A clip defaults to 241 frames, about 10 seconds. Every GPU employee masses the total 29B mannequin in BF16. IFM publishes no VRAM determine or era time.

Prompts work greatest as structured captions. IFM ships a Conceptual Reasoner ability for Claude Code and Codex that rewrites free textual content. Captions have a tough price range of 1,536 tokens.

IFM strongly recommends the separate pan2-guardrail. In IFM exams, it minimize unsafe outputs from dangerous T2V requests from 33.9% to 1.5%. It additionally blocked 11.5% of benign T2V requests.