NVIDIA Releases Nemotron 3.5 Tremendous VL, a 120B Open Multimodal Agent Mannequin

NVIDIA simply launched Nemotron 3.5 Tremendous VL, an open vision-language mannequin constructed for always-on enterprise brokers. Nemotron 3.5 Tremendous VL is a 120B-parameter mixture-of-experts mannequin from NVIDIA that prompts 12B parameters per token. It reads textual content, pictures, paperwork and video, then causes, calls instruments and writes code in a single mannequin. For builders, 1 open checkpoint replaces a separate textual content mannequin together with a separate imaginative and prescient mannequin. NVIDIA additionally shipped 2 edge fashions, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B.

TL;DR

Dimension: 120B complete parameters, 12B energetic (MoE). Context as much as 1M tokens.
Runs on: 1x DGX Spark, 1x H200, 2x H100 or 1x GB200 with the NVFP4 checkpoint. A BF16 checkpoint can also be printed.
Efficiency: Robust reasoning, agentic coding and doc imaginative and prescient. NVIDIA claims as much as 5x greater single-GPU throughput than open friends.
Finest: 89.90 on GPQA-Diamond (NVFP4).
Worst: 17.78 on tau3-bench banking (NVFP4).

What’s Nemotron 3.5 Tremendous VL?

It’s an open multimodal agent mannequin. Weights ship beneath the OpenMDW-1.1 license. NVIDIA additionally launched open datasets, coaching recipes and RL environments in NeMo Gymnasium.

The mannequin card lists textual content, picture and video inputs with textual content output. Video is restricted to MP4 clips as much as 2 minutes. Reasoning could be toggled on or off via the chat template.

How does the structure work?

The spine interleaves Mamba-2, mixture-of-experts and a focus layers. Every MoE layer holds 512 routed consultants. A C-RADIO imaginative and prescient encoder and a connector feed pictures, doc pages and video frames into the language spine.

Velocity comes from 4 methods:

Multi-token prediction: a local drafter proposes a number of tokens, verified in 1 ahead go.
LiLiCorr: a speculative drafter for DGX Spark, as much as 2.3x quicker token technology.
NVFP4 with quantization-aware distillation: 4-bit weights and activations on Blackwell, 4-bit weights with 16-bit activations elsewhere.
Context reuse: prefix caching retains each consideration caches and Mamba states throughout agent turns.

The mannequin was distilled from Nemotron 3 Extremely. NVIDIA crew states that it skilled on about 26 trillion tokens. Put up-training used roughly 3.8 million samples plus RL throughout dozens of environments. Continued pretraining added 10B Indic tokens.

How does it carry out on benchmarks?

All scores under are NVIDIA-reported, from the NVFP4 checkpoint on the mannequin card.

Space	Benchmark	Rating
Reasoning	GPQA-Diamond	89.90
Coding	SWE-bench Verified	71.07
Coding	Terminal-Bench 2.1	60.80
Net agent	BrowseComp	77.5
Instruction following	IFBench	71.50
Charts	CharXiv reasoning	79.88
Paperwork	MMLongBench-Doc	62.82
Laptop use	OSWorld-Verified	52.3
Lengthy-horizon agent	tau3-bench banking	17.78

On agent harness evals, Tremendous VL beats Nemotron 3 Extremely. It scores 63.44% on Hermes Bench versus 56.33%. On LangChain evals it scores 87.05% versus 84.3%. Extremely carries about 4.6x extra complete parameters.

NVIDIA crew studies over 600 output tokens per second on Synthetic Evaluation agentic duties by way of a Baseten endpoint. Its chart marks a 5x pace hole over Inkling Small.

<div class="panel on" id="p-ov">
<div class="grid">
<div class="stat"><b data-n="120">0</b><span>B complete parameters</span></div>
<div class="stat"><b data-n="12">0</b><span>B energetic per token</span></div>
<div class="stat"><b data-n="1" data-s="M">0</b><span>token context</span></div>
<div class="stat"><b data-n="512">0</b><span>routed consultants per MoE layer</span></div>
</div>
<div class="moe">
<div class="lbl"><span>MoE routing: every token lights up a small slice of the community</span><span id="mtp-tok">token #1</span></div>
<div class="cells" id="mtp-cells"></div>
</div>
<p class="observe"><robust>What it’s:</robust> an open vision-language mannequin for always-on brokers. It reads textual content, pictures, paperwork and video, then causes, calls instruments and writes code. Weights ship beneath OpenMDW-1.1 with open datasets and recipes.</p>
</div>

<div class="panel" id="p-hw2">
<div class="pipe" id="mtp-pipe">
<div class="stage on" data-i="0">Inputs<small>textual content, picture, video</small></div><div class="arr"></div>
<div class="stage" data-i="1">C-RADIO<small>imaginative and prescient encoder</small></div><div class="arr"></div>
<div class="stage" data-i="2">Hybrid MoE<small>Mamba-2 + consideration</small></div><div class="arr"></div>
<div class="stage" data-i="3">MTP decode<small>+ LiLiCorr</small></div><div class="arr"></div>
<div class="stage" data-i="4">Output<small>textual content, device calls</small></div>
</div>
<div class="element" id="mtp-detail"></div>
<div class="ctrl">
<button class="btn" id="mtp-prev"> Again</button>
<button class="btn pri" id="mtp-next">Subsequent </button>
<button class="btn" id="mtp-play"> Auto-play</button>
<span class="step" id="mtp-step">Step 1 of 5</span>
</div>
</div>

<div class="panel" id="p-bm">
<div class="seg"><button class="on" data-c="nv">NVFP4</button><button data-c="bf">BF16</button></div>
<div id="mtp-bars"></div>
<p class="observe">NVIDIA-reported scores from the mannequin card. Orange bar marks the weakest outcome.</p>
</div>

<div class="panel" id="p-cp">
<div class="cmp" id="mtp-cmp"></div>
<p class="legend">Vendor self-reported scores. Tremendous VL makes use of NVFP4. Harnesses differ: Inkling-Small SWE-bench makes use of a bash-only harness.</p>
</div>

<div class="panel" id="p-dp">
<div class="hw" id="mtp-hw">
<button class="chip on" data-h="0">1x DGX Spark<small>vLLM · LiLiCorr</small></button>
<button class="chip" data-h="1">1x H200<small>vLLM · MTP 3</small></button>
<button class="chip" data-h="2">2x H100<small>vLLM · MTP 3</small></button>
<button class="chip" data-h="3">1x GB200<small>vLLM · MTP 5</small></button>
</div>
<div class="element" id="mtp-hwd"></div>
<pre id="mtp-code"></pre>
<p class="observe">Instructions abridged from the Hugging Face mannequin card. Use the total flag set there for manufacturing.</p>
</div>

<div class="ft"><span>Sources: NVIDIA weblog, NVIDIA technical weblog, Hugging Face mannequin playing cards</span><span class="br">© Marktechpost</span></div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-nsvl’);
perform $(s){return R.querySelector(s)}
perform $$(s){return Array.prototype.slice.name(R.querySelectorAll(s))}
perform resize(){attempt{guardian.postMessage({mtpEmbed:’nsvl’,h:R.offsetHeight+40},’*’)}catch(e){}}
/* tabs */
$$(‘.tab’).forEach(perform(t){t.addEventListener(‘click on’,perform(){
$$(‘.tab’).forEach(perform(x){x.classList.take away(‘on’)});t.classList.add(‘on’);
$$(‘.panel’).forEach(perform(p){p.classList.take away(‘on’)});
var ok=t.getAttribute(‘data-t’);$(‘#p-‘+ok).classList.add(‘on’);
if(ok===’bm’)drawBars(cur);if(ok===’cp’)drawCmp();if(ok===’ov’)depend();
setTimeout(resize,60);
})});
/* counters */
perform depend(){$$(‘.stat b’).forEach(perform(b){var n=+b.getAttribute(‘data-n’),s=b.getAttribute(‘data-s’)||”,t0=null;
perform f(ts){if(!t0)t0=ts;var p=Math.min((ts-t0)/900,1),e=1-Math.pow(1-p,3);b.textContent=Math.spherical(n*e)+s;if(p<1)requestAnimationFrame(f)}requestAnimationFrame(f)})}
depend();
/* MoE cells */
var cells=$(‘#mtp-cells’),N=120;for(var i=0;i<N;i++){var c=doc.createElement(‘div’);c.className=”cell”;cells.appendChild(c)}
var cs=$$(‘.cell’),tk=1;
perform route(){cs.forEach(perform(c){c.classList.take away(‘act’)});var used={};for(var ok=0;ok<12;ok++){var j;do{j=Math.flooring(Math.random()*N)}whereas(used[j]);used[j]=1;cs[j].classList.add(‘act’)}$(‘#mtp-tok’).textContent=”token #”+(tk++)+’ · 12 of 120 cells energetic’}
route();setInterval(route,1100);
/* pipeline */
var D=[
‘<b>Inputs.</b> Text, RGB images (jpeg, png) and MP4 video up to 2 minutes. Documents can be passed as page images with text. Everything shares a context window of up to 1M tokens.’,
‘<b>C-RADIO vision encoder + connector.</b> Turns images, document pages and video frames into representations the language backbone can process. These parts stay in BF16 after quantization.’,
‘<b>Hybrid MoE backbone.</b> Interleaved Mamba-2, MoE and attention layers. 512 routed experts per MoE layer, 12B of 120B parameters active per token. Distilled from Nemotron 3 Ultra.’,
‘<b>Fast decoding.</b> Native multi-token prediction drafts several tokens for 1-pass verification. On DGX Spark, the LiLiCorr drafter gives up to 2.3x faster token generation.’,
‘<b>Output.</b> Text answers, reasoning (toggle with enable_thinking) and tool calls. NVIDIA reports 600+ output tokens per second on Artificial Analysis agentic tasks via Baseten.’];
var si=0,timer=null,st=$$(‘.stage’);
perform present(i){si=(i+5)%5;st.forEach(perform(s,ok){s.classList.toggle(‘on’,ok===si)});$(‘#mtp-detail’).innerHTML=D[si];$(‘#mtp-step’).textContent=”Step “+(si+1)+’ of 5′;setTimeout(resize,30)}
st.forEach(perform(s){s.addEventListener(‘click on’,perform(){present(+s.getAttribute(‘data-i’))})});
$(‘#mtp-next’).addEventListener(‘click on’,perform(){present(si+1)});
$(‘#mtp-prev’).addEventListener(‘click on’,perform(){present(si-1)});
$(‘#mtp-play’).addEventListener(‘click on’,perform(){var b=this;if(timer){clearInterval(timer);timer=null;b.textContent=” Auto-play”}else{timer=setInterval(perform(){present(si+1)},2200);b.textContent=”❚❚ Pause”}});
present(0);
/* benchmarks */
var B=[[‘GPQA-Diamond’,89.90,89.52],[‘CharXiv reasoning’,79.88,80.82],[‘BrowseComp’,77.5,78],[‘SWE-bench Verified’,71.07,71.10],[‘IFBench’,71.50,70.86],[‘MMLongBench-Doc’,62.82,62.82],[‘Terminal-Bench 2.1’,60.80,66.70],[‘OSWorld-Verified’,52.3,51.14],[‘tau3-bench banking’,17.78,18.88]];
var cur=”nv”;
perform drawBars(c){cur=c;var h=””;B.forEach(perform(r,i){var v=c===’nv’?r[1]:r[2];h+='<div class="bar"><span class="nm">’+r[0]+'</span><div class="monitor"><div class="fill’+(i===B.length-1?’ low’:”)+’" data-w="’+v+’"></div></div><span class="val">’+v.toFixed(2)+'</span></div>’});
$(‘#mtp-bars’).innerHTML=h;requestAnimationFrame(perform(){requestAnimationFrame(perform(){$$(‘#mtp-bars .fill’).forEach(perform(f){f.type.width=f.getAttribute(‘data-w’)+’%’})})});setTimeout(resize,60)}
$$(‘.seg button’).forEach(perform(b){b.addEventListener(‘click on’,perform(){$$(‘.seg button’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawBars(b.getAttribute(‘data-c’))})});
/* evaluate */
var C=[[‘GPQA Diamond’,’89.90′,’86.6′,’89.5′],[‘SWE-bench Verified’,’71.07′,’72.0′,’80.2′],[‘IFBench’,’71.50′,’76.1′,’82.2′],[‘CharXiv reasoning’,’79.88′,’77.2′,’77.4′],[‘BrowseComp’,’77.5′,’63.8′,’77.4′]];
perform drawCmp(){var h=””;C.forEach(perform(r){h+='<div class="row"><div class="rt">’+r[0]+'</div>’+
‘<div class="b me"><span class="nm">Tremendous VL</span><div class="monitor"><div class="fill" data-w="’+r[1]+’"></div></div><span class="val">’+r[1]+'</span></div>’+
‘<div class="b"><span class="nm">Qwen3.5-122B-A10B</span><div class="monitor"><div class="fill q" data-w="’+r[2]+’"></div></div><span class="val">’+r[2]+'</span></div>’+
‘<div class="b"><span class="nm">Inkling-Small</span><div class="monitor"><div class="fill i" data-w="’+r[3]+’"></div></div><span class="val">’+r[3]+'</span></div></div>’});
$(‘#mtp-cmp’).innerHTML=h;requestAnimationFrame(perform(){requestAnimationFrame(perform(){$$(‘#mtp-cmp .fill’).forEach(perform(f){f.type.width=f.getAttribute(‘data-w’)+’%’})})});setTimeout(resize,60)}
/* deploy */
var H=[
[‘<b>1x DGX Spark (GB10).</b> Text and image, 1M-token context in vLLM. NVIDIA recommends the LiLiCorr drafter for all Spark deployments.’,’vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 n –max-num-seqs 1 –trust-remote-code n –speculative_config.model nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr n –speculative_config.num_speculative_tokens 7 n –speculative_config.method dflash n –reasoning-parser nemotron_v3′],
[‘<b>1x H200.</b> Text, image and video, 1M-token context, MTP with 3 draft tokens. Same command as 2x H100 with tensor parallel 1.’,’vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 n –tensor-parallel-size 1 –trust-remote-code n –speculative-config.method mtp n –speculative-config.num_speculative_tokens 3 n –reasoning-parser nemotron_v3′],
[‘<b>2x H100.</b> Text, image and video, 1M-token context, MTP with 3 draft tokens. Runs W4A16 on non-Blackwell GPUs.’,’vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 n –tensor-parallel-size 2 –trust-remote-code n –speculative-config.method mtp n –speculative-config.num_speculative_tokens 3 n –reasoning-parser nemotron_v3′],
[‘<b>1x GB200.</b> Text, image and video, 1M-token context, MTP with 5 draft tokens. Full W4A4 NVFP4 on Blackwell. SGLang and TensorRT-LLM recipes also exist.’,’vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 n –max-num-seqs 512 –trust-remote-code n –speculative_config.method mtp n –speculative_config.num_speculative_tokens 5 n –reasoning-parser nemotron_v3′]];
perform hw(i){$$(‘.chip’).forEach(perform(c,ok){c.classList.toggle(‘on’,ok===i)});$(‘#mtp-hwd’).innerHTML=H[i][0];$(‘#mtp-code’).textContent=H[i][1];setTimeout(resize,30)}
$$(‘.chip’).forEach(perform(c){c.addEventListener(‘click on’,perform(){hw(+c.getAttribute(‘data-h’))})});hw(0);
window.addEventListener(‘load’,resize);window.addEventListener(‘resize’,resize);setTimeout(resize,300);
})();
</script>
</physique>
</html>
“>

(perform(){var f=doc.getElementById(‘mtp-nsvl-frame’);
window.addEventListener(‘message’,perform(e){var d=e.information;if(d&&d.mtpEmbed===’nsvl’&&f&&e.supply===f.contentWindow&&d.h>200){f.type.top=d.h+’px’;}});})();

How does it evaluate with related open fashions?

Characteristic	Nemotron 3.5 Tremendous VL	Qwen3.5-122B-A10B	Inkling-Small	Nemotron 3 Extremely
Developer	NVIDIA	Alibaba Qwen	Pondering Machines Lab	NVIDIA
Whole / energetic params	120B / 12B	122B / 10B	276B / 12B	550B / 55B
Context	As much as 1M	262,144 native, extensible to 1,010,000	Not disclosed on mannequin card	Evaluated at 1M (Ruler @1M)
Inputs	Textual content, picture, video	Textual content, picture, video	Textual content, picture, audio	Not disclosed
License	OpenMDW-1.1	Apache 2.0	Apache 2.0	OpenMDW-1.1
Reference {hardware}	1x DGX Spark, 1x H200, 2x H100 or 1x GB200	8-GPU tensor-parallel serving instance	Not disclosed	Not disclosed
GPQA Diamond	89.90	86.6	89.5	Not disclosed
SWE-bench Verified	71.07	72.0	80.2 (bash-only harness)	65 to 70.4 (throughout 5 harnesses)
IFBench	71.50	76.1	82.2	82
CharXiv reasoning	79.88	77.2	77.4	Not disclosed
BrowseComp	77.5	63.8	77.4 (with context administration)	Not disclosed
Terminal-Bench	60.80 (v2.1)	49.4 (v2)	64.7 (v2.1, greatest harness)	54 (v2.0)

Scores are self-reported by every vendor. Harnesses and benchmark variations differ, so deal with gaps of some factors with warning.

How can builders deploy it?

Tremendous VL is on the market on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM. Supported runtimes are vLLM, SGLang and TensorRT-LLM on Blackwell and Hopper. Native paths embrace llama.cpp and Unsloth.

For manufacturing, Dynamo recipes add KV-aware routing. On 16 B200 GPUs at 50 tokens per second per person, NVIDIA studies 8x extra customers served. First-token latency drops 85%.

Customization runs via NeMo AutoModel for SFT and LoRA, and NeMo RL for reinforcement studying. NVIDIA cites early outcomes:

Scale AI for Skydance vendor credit score processing: 40x decrease inference price and 5x decrease latency.
Ramp receipt-to-JSON extraction: as much as 50% greater throughput and 10x decrease latency anticipated.
AI21 for Alice risk intelligence: 40% greater accuracy.

What are the Nemotron 3.5 Bolt fashions?

Nemotron 3.5 Bolt 74B-A9B targets native brokers. NVIDIA’s launch notes describe it as distilled from Tremendous VL. Native 2-bit quantization matches it into 24GB of reminiscence on RTX and DGX Spark techniques. NVIDIA says it ranks #1 amongst open fashions in its class on PinchBench. It ranks #2 on BrowseComp amongst open fashions. It’s accessible now on Hugging Face with llama.cpp, Ollama, LM Studio, Unsloth and vLLM help.

Bolt E4B 4GB arrives October 23. It targets voice assistants and embedded brokers on GeForce RTX, Jetson and RTX or DGX Spark.

Key Takeaways

120B MoE, 12B energetic, 1M context, open weights beneath OpenMDW-1.1.
1 mannequin handles textual content, pictures, paperwork, video, instruments and code.
NVFP4 checkpoint runs on a single DGX Spark, H200 or GB200.
89.90 GPQA-Diamond, however solely 17.78 on tau3-bench banking.
Bolt 74B-A9B brings a 24GB native agent mannequin to RTX PCs.

Take a look at the technical report, mannequin card and utilization cookbook. All credit score goes to the researcher of this challenge. Additionally, be at liberty to comply with us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you may be part of us on telegram as effectively.

The submit NVIDIA Releases Nemotron 3.5 Tremendous VL, a 120B Open Multimodal Agent Mannequin appeared first on MarkTechPost.