Meet Drex Max: Nace AI’s Multimodal Determination Mannequin for Laptop-Use Brokers with a Rating 67.59

Nace AI has launched Drex Max, a multimodal resolution mannequin for brokers that use computer systems and browsers. The Drex Max multimodal resolution mannequin reads textual content and pictures and returns calibrated chances. It writes no textual content. Nace staff experiences 67.59 on the general public Determination Index 0.3.2, 6.9 factors above TypeSafe’s Jev 1.14. For agent builders, meaning one low cost, typed name per click on.

TL;DR

Dimension: Parameters not disclosed. As much as 262,144 tokens per query on the API, together with picture tokens.
Runs on: Hosted API solely (drex-max on Nace Console). No open weights or quantized builds introduced.
Efficiency: 67.59 on Determination Index 0.3.2 (public), the highest rating Nace lists for any resolution mannequin.
Greatest: 98.2 on BFCL instrument calling, and 92.2 on JevBench versus 86.7 for Jev 1.14.
Worst: 19.0 on HLE, and 72.5 on MMLU-Professional versus 82.1 for Jev 1.14.

What’s Drex Max?

Drex Max is a call mannequin from Nace AI that solutions typed questions on textual content and pictures. It’s the most correct mannequin within the Drex household, and the one one which reads pictures.

A choice mannequin doesn’t generate tokens. You ship a state and named questions. Every query is one in all 3 sorts: noul (sure/no), alternative (choose a label) or rating (ordered scale). The mannequin returns a chance distribution for every reply in 1 HTTP name. Drex can solely reply with the choices you gave it.

How does Drex Max work with pictures?

Drex Max makes use of the identical POST /v1/systemone endpoint as each Drex mannequin. You choose it with "mannequin": "drex-max". It’s a pinned ID, so drex-latest doesn’t path to it.

Photos go in an pictures array. Each query sees the identical pictures, earlier than the state and questions. You check with them as ‘Picture 1’ and ‘Picture 2’ in plain textual content. Every entry is both an uploaded workspace file or inline base64.

Key picture guidelines from the Drex Max information:

Codecs: nonetheless PNG, JPEG or WebP.
As much as 8 pictures per request, 16 MiB every, 32 MiB whole.
Every picture is resized to at most 1,638,400 pixels (1,280 × 1,280).
Every picture prices as much as 1,602 tokens in each query.
As much as 16 questions per request, with a 4 MiB request physique.

How does Drex Max carry out on benchmarks?

All scores under come from Nace’s Drex web page. They’re self-reported. Nace says Drex trains on the index’s official coaching splits and is scored on held-out splits.

Determination benchmarks: Drex Max scores 67.6 on the general public Determination Index 0.3.2. Jev 1.14 scores 60.7 and Drex 2.0 scores 61.3. On JevBench, Drex Max reaches 92.2 in opposition to 86.7 for Jev 1.14.
Agentic benchmarks: Drex Max posts 0.780 on OSWorld-Verified and 98.2 on BFCL. It scores 85.2 on When2Call and 84.8 on RAGTruth.
Imaginative and prescient: On the Determination Index Imaginative and prescient public suite, Drex Max scores 74.5. GPT-6 Luna Choices scores 63.9. Mind2Web reaches 88.6 and InfographicVQA reaches 98.8.
The place it trails: Jev 1.14 leads on MMLU-Professional (82.1 vs 72.5), HLE (19.4 vs 19.0) and ARC-Problem (97.6 vs 97.4). On Hateful Memes, Drex Max trails GPT-6 Luna Choices, 70.6 vs 71.2.

<div class="tabs" function="tablist">
<button class="tab" function="tab" aria-selected="true" data-t="movement">The way it works</button>
<button class="tab" function="tab" aria-selected="false" data-t="bench">Benchmarks</button>
<button class="tab" function="tab" aria-selected="false" data-t="vs">vs Jev 1.14</button>
<button class="tab" function="tab" aria-selected="false" data-t="lim">API limits</button>
<button class="tab" function="tab" aria-selected="false" data-t="value">Value calculator</button>
</div>

<div class="pn on" data-p="movement">
<div class="movement">
<div class="node" data-n="0"><h4>1 · Enter</h4>
<p><code>state</code>: textual content or JSON</p>
<p><code>pictures</code>: as much as 8, nonetheless PNG / JPEG / WebP</p>
<div class="imgs"><span></span><span></span><span></span></div>
</div>
<div class="arr"><i></i></div>
<div class="node" data-n="1"><h4>2 · Questions</h4>
<p><code>noul</code> sure / no</p>
<p><code>alternative</code> choose a label</p>
<p><code>rating</code> ordered scale</p>
<p model="coloration:#a8a29e">As much as 16 per request</p>
</div>
<div class="arr"><i></i></div>
<div class="node" data-n="2"><h4>3 · Reply</h4>
<p model="margin-bottom:4px">"Which button subsequent?"</p>
<div class="prob"><em>search</em><div class="b"><u data-w="81"></u></div><b>0.81</b></div>
<div class="prob"><em>cart</em><div class="b"><u data-w="12"></u></div><b>0.12</b></div>
<div class="prob"><em>again</em><div class="b"><u data-w="7"></u></div><b>0.07</b></div>
</div>
</div>
<p class="cap" id="mtp-cap"></p>
<div class="ctl">
<button class="btn" id="mtp-prev"> Again</button>
<button class="btn pri" id="mtp-next">Subsequent step </button>
<button class="btn" id="mtp-play">Auto play</button>
<div class="dots"><i></i><i></i><i></i></div>
</div>
<p class="notice">Reply chances are an illustration of the response form, not actual mannequin output.</p>
</div>

<div class="pn" data-p="bench">
<div class="seg" id="mtp-seg">
<button class="on" data-s="di">Determination Index</button>
<button data-s="ag">Agentic</button>
<button data-s="vi">Imaginative and prescient</button>
</div>
<div id="mtp-bars"></div>
<p class="notice" id="mtp-bnote"></p>
</div>

<div class="pn" data-p="vs">
<desk class="vs">
<thead><tr><th>Benchmark</th><th model="text-align:proper">Drex Max</th><th model="text-align:proper">Jev 1.14</th><th>Outcome</th></tr></thead>
<tbody id="mtp-vs"></tbody>
</desk>
<p class="notice">All figures from Nace’s Drex web page (self-reported). Jev 1.14 rejects 15.9% of screenshots by measurement, per Nace.</p>
</div>

<div class="pn" data-p="lim">
<div class="grid" id="mtp-lims"></div>
<p class="notice">Faucet a card for element. Supply: Nace Drex Max API information.</p>
</div>

<div class="pn" data-p="value">
<div class="calc">
<div>
<label for="mtp-dec">Choices per day: <b id="mtp-decv" model="coloration:#f5f5f4">10,000</b></label>
<enter kind="vary" id="mtp-dec" min="1000" max="1000000" step="1000" worth="10000">
<label for="mtp-img">Screenshots per resolution: <b id="mtp-imgv" model="coloration:#f5f5f4">1</b></label>
<enter kind="vary" id="mtp-img" min="0" max="8" step="1" worth="1">
<label for="mtp-txt">Textual content tokens per resolution (state + query): <b id="mtp-txtv" model="coloration:#f5f5f4">500</b></label>
<enter kind="vary" id="mtp-txt" min="50" max="20000" step="50" worth="500">
</div>
<div class="out">
<small>Estimated day by day value, higher certain</small>
<div class="large" id="mtp-cost">$0.00</div>
<small id="mtp-tok"></small>
<small model="margin-top:10px">Makes use of $0.06 per 1M enter tokens and the 1,602-token most per picture. Output tokens usually are not billed. Actual picture token counts will be decrease.</small>
</div>
</div>
</div>

<div class="ft">
<span class="src">Sources: <a href="https://www.nace.ai/drex" goal="_blank" rel="noopener">nace.ai/drex</a> · <a href="https://console.nace.ai/docs/drex/guides/drex-max" goal="_blank" rel="noopener">Drex Max information</a></span>
<a href="https://www.marktechpost.com" goal="_blank" rel="noopener">© Marktechpost</a>
</div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-drexmax’);
perform $(s,c)perform $$(s,c)R).querySelectorAll(s))
perform publish(){attempt{var h=R.offsetHeight+40;mother or father.postMessage({mtpDrexMaxHeight:h,kind:’mtp-resize’,top:h},’*’)}catch(e){}}
/* counters */
$$(‘[data-count]’).forEach(perform(el){var to=+el.dataset.depend,d=+(el.dataset.dec||0),suf=el.dataset.suf||”,sep=el.dataset.sep,t0=null;
perform f(v){var s=v.toFixed(d);if(sep)s=Math.spherical(v).toLocaleString(‘en-US’);return s+suf}
perform step(ts){if(!t0)t0=ts;var okay=Math.min(1,(ts-t0)/1100),e=1-Math.pow(1-k,3);el.textContent=f(to*e);if(okay<1)requestAnimationFrame(step);else el.textContent=f(to)}
requestAnimationFrame(step)});
/* tabs */
$$(‘.tab’).forEach(perform(b){b.addEventListener(‘click on’,perform(){$$(‘.tab’).forEach(perform(x){x.setAttribute(‘aria-selected’,x===b)});
$$(‘.pn’).forEach(perform(p){p.classList.toggle(‘on’,p.dataset.p===b.dataset.t)});
if(b.dataset.t===’bench’)drawBars(cur);if(b.dataset.t===’movement’)present(stp);setTimeout(publish,60);setTimeout(publish,500)})});
/* movement */
var caps=[‘Step 1: send a state plus up to 8 still images. Every question sees the same images, before the state and questions.’,’Step 2: name up to 16 typed questions. Drex Max can only answer with the options you supply.’,’Step 3: one call returns a probability for every option. Nace lists 174 ms median per screenshot decision.’];
var stp=0,timer=null;
perform present(i){stp=i;$$(‘.node’).forEach(perform(n){n.classList.toggle(‘lit’,+n.dataset.n<=i)});
$$(‘.dots i’).forEach(perform(d,okay){d.classList.toggle(‘on’,okay===i)});$(‘#mtp-cap’).textContent=caps[i];
$$(‘.prob u’).forEach(perform(u){u.model.width=i===2?u.dataset.w+’%’:’0′})}
$(‘#mtp-next’).onclick=perform(){stopPlay();present((stp+1)%3)};$(‘#mtp-prev’).onclick=perform(){stopPlay();present((stp+2)%3)};
perform stopPlay(){if(timer){clearInterval(timer);timer=null;$(‘#mtp-play’).textContent=”Auto play”}}
$(‘#mtp-play’).onclick=perform(){if(timer){stopPlay();return}this.textContent=”Pause”;present(0);timer=setInterval(perform(){present((stp+1)%3)},2200)};
present(0);
/* bars */
var D={di:{max:80,rows:[[‘Drex Max’,67.6,1],[‘Drex 2.0’,61.3],[‘Jev 1.14’,60.7],[‘Drex 1.5’,58.1],[‘GPT-6 Luna Decisions’,47.8]],notice:’Determination Index 0.3.2, public benchmarks solely, chance-corrected. Supply: nace.ai/drex (self-reported).’},
ag:{max:100,rows:[[‘Drex Max · BFCL’,98.2,1],[‘Jev 1.14 · BFCL’,96.4],[‘Drex Max · OSWorld-Verified’,78.0,1],[‘Jev 1.14 · OSWorld-Verified’,73.4],[‘Drex Max · When2Call’,85.2,1],[‘Jev 1.14 · When2Call’,81.7]],notice:’OSWorld-Verified proven ×100 (0.780 and 0.734). Supply: nace.ai/drex (self-reported).’},
vi:{max:100,rows:[[‘Drex Max · Vision Index’,74.5,1],[‘GPT-6 Luna Decisions’,63.9],[‘Drex Max · Mind2Web’,88.6,1],[‘GPT-6 Luna · Mind2Web’,81.2],[‘Drex Max · CV-Bench’,86.3,1],[‘GPT-6 Luna · CV-Bench’,73.8]],notice:’Determination Index Imaginative and prescient, public suite, 11 benchmarks. Supply: nace.ai/drex (self-reported).’}};
var cur=”di”;
perform drawBars(okay){cur=okay;var d=D[k],h=””;d.rows.forEach(perform(r){h+='<div class="row’+(r[2]?’ me’:”)+’"><span class="lab">’+r[0]+'</span><div class="tr"><div class="fl" data-w="’+(r[1]/d.max*100)+’"></div></div><span class="v">’+r[1].toFixed(1)+'</span></div>’});
$(‘#mtp-bars’).innerHTML=h;$(‘#mtp-bnote’).textContent=d.notice;
requestAnimationFrame(perform(){requestAnimationFrame(perform(){$$(‘#mtp-bars .fl’).forEach(perform(f){f.model.width=f.dataset.w+’%’})})});
$$(‘#mtp-seg button’).forEach(perform(b){b.classList.toggle(‘on’,b.dataset.s===okay)});setTimeout(publish,60)}
$$(‘#mtp-seg button’).forEach(perform(b){b.onclick=perform(){drawBars(b.dataset.s)}});
/* vs */
var V=[[‘Decision Index 0.3.2′,’67.6′,’60.7’,1],[‘JevBench’,’92.2′,’86.7′,1],[‘Decision Bench’,’94.3′,’92.3′,1],[‘Game Bench win %’,’64.0′,’50.0′,1],[‘BFCL’,’98.2′,’96.4′,1],[‘OSWorld-Verified’,’0.780′,’0.734′,1],[‘Screenshot decision’,’174 ms’,’253 ms’,1],[‘MMLU-Pro’,’72.5′,’82.1′,0],[‘HLE’,’19.0′,’19.4′,0],[‘ARC-Challenge’,’97.4′,’97.6′,0]];
$(‘#mtp-vs’).innerHTML=V.map(perform(r){return ‘<tr><td>’+r[0]+'</td><td class="n">’+r[1]+'</td><td class="n">’+r[2]+'</td><td>’+(r[3]?'<span class="capsule w">Max leads</span>’:'<span class="capsule l">Jev leads</span>’)+'</td></tr>’}).be part of(”);
/* limits */
var L=[[’16’,’questions per request’,’Same limit with or without images.’],[‘262,144′,’tokens per question’,’State plus that question, image tokens included. Nace’s landing page lists 131K context.’],[‘8′,’images per request’,’Still PNG, JPEG or WebP. 16 MiB each, 32 MiB total.’],[‘1,602′,’max tokens per image’,’Charged in every question. Images resize to at most 1,638,400 pixels (1,280 × 1,280).’],[‘4 MiB’,’request body’,’Inline base64 images count toward it, so about 3 MiB of image bytes fit.’],[‘$0′,’for failed requests’,’Only a 200 response is billed. Output tokens are never billed.’]];
$(‘#mtp-lims’).innerHTML=L.map(perform(l){return ‘<div class="lim" tabindex="0"><b>’+l[0]+'</b><span>’+l[1]+'</span><p>’+l[2]+'</p></div>’}).be part of(”);
$$(‘.lim’).forEach(perform(c){perform t(){c.classList.toggle(‘open’);setTimeout(publish,380)}c.onclick=t;c.onkeydown=perform(e){if(e.key===’Enter’||e.key===’ ‘){e.preventDefault();t()}}});
/* calc */
perform calc(){var n=+$(‘#mtp-dec’).worth,im=+$(‘#mtp-img’).worth,tx=+$(‘#mtp-txt’).worth,tok=n*(tx+im*1602),c=tok/1e6*0.06;
$(‘#mtp-decv’).textContent=n.toLocaleString(‘en-US’);$(‘#mtp-imgv’).textContent=im;$(‘#mtp-txtv’).textContent=tx.toLocaleString(‘en-US’);
$(‘#mtp-cost’).textContent=”$”+(c<10?c.toFixed(2):Math.spherical(c).toLocaleString(‘en-US’));$(‘#mtp-tok’).textContent=tok.toLocaleString(‘en-US’)+’ enter tokens per day’}
[‘#mtp-dec’,’#mtp-img’,’#mtp-txt’].forEach(perform(s){$(s).oninput=calc});calc();
window.addEventListener(‘load’,publish);window.addEventListener(‘resize’,publish);setTimeout(publish,300);setTimeout(publish,1200);
})();
</script>
</physique></html>
” model=”width:100%;top:600px;border:0;show:block;background:clear;” loading=”lazy” scrolling=”no”>

(perform(){var f=doc.getElementById(‘mtp-drexmax-frame’);window.addEventListener(‘message’,perform(e)e.supply!==f.contentWindow)return;var h=e.knowledge.mtpDrexMaxHeight;if(h&&h>200)f.model.top=h+’px’;);})();

How briskly and the way low cost is Drex Max?

Nace lists a 174 ms median per screenshot resolution. Jev 1.14 takes 253 ms on the identical 243 questions. Nace says Drex Max accepts each screenshot measurement, whereas Jev 1.14 rejects 15.9% by measurement.

Throughout its benchmark desk, Nace experiences a 183 ms median latency. The launch publish cites 34 ms per resolution and beneath 1 second at 262K tokens. These figures use completely different setups, so deal with them as separate claims.

Pricing is $0.06 per 1M enter tokens. Output tokens usually are not billed. New accounts get $25 in signup credit score, and failed requests value nothing.

How does Drex Max examine with different resolution fashions?

Characteristic	Drex Max	Jev 1.14	Drex 2.0	d3-flash
Developer	Nace AI	TypeSafe AI	Nace AI	vLLM Semantic Router
Parameters	Not disclosed	Not disclosed	35B whole, about 3B lively (MoE)	8.39B
Context	262,144 tokens per query	64K per request, 262K with long-context header	16,384 default, as much as 131,072	Not disclosed
Inputs	Textual content, pictures	Textual content (TypeSafe docs)	Textual content	Textual content, pictures, video
License	Proprietary (hosted)	Proprietary (hosted)	Nace.AI Open RAIL-M	Apache-2.0
Determination Index (public)	67.6 (0.3.2)	60.7 (0.3.2, per Nace)	61.25 (0.3.1)	57.8 (0.3.2)
Imaginative and prescient index (public)	74.5	Not disclosed	Not relevant	69.9
Worth per 1M enter tokens	$0.06	$0.042	$0.06	Self-host, free trial
Runs on	Nace Console API	TypeSafe API	Single accelerator, bf16 about 70 GB	1 AMD Intuition MI325X examined

Word: d3-flash’s imaginative and prescient determine is its personal analysis on the 0.3.2 imaginative and prescient board. Drex Max’s determine is Nace’s personal run. Cross-vendor numbers usually are not a managed comparability.

Key Takeaways

Drex Max is Nace AI’s first multimodal resolution mannequin.
It scores 67.59 on Determination Index 0.3.2, 6.9 above Jev 1.14.
Screenshot choices take 174 ms, versus 253 ms for Jev 1.14.
Pricing is $0.06 per 1M enter tokens, with output free.
Dimension is undisclosed, weights are closed, and scores are self-reported.

Take a look at the Drex Max information, the Drex product web page and the launch publish. Additionally learn our earlier protection of Drex 1.5. All credit score goes to the researcher of this venture. Additionally, be at liberty to comply with us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you’ll be able to be part of us on telegram as effectively.