NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Mannequin for Native Voice Brokers, Sport Characters and Robots

NVIDIA has simply launched Nemotron 3.5 Bolt E4B 4GB, an open small language mannequin for quick conversational brokers. It targets voice assistants, recreation NPCs and robotic interfaces working domestically on RTX, Jetson and Spark {hardware}. The mannequin is post-trained from Google’s Gemma 4 E4B utilizing reinforcement studying solely. For builders, that is agent habits in about 4GB of VRAM. That leaves room for speech and imaginative and prescient fashions.

TL;DR

Measurement: 8B whole parameters, 4B efficient, 64K context.
Runs on: GeForce RTX, Jetson, RTX Spark and DGX Spark. The Q4_K_M construct used 3.1 to 4.3 GB VRAM on RTX in NVIDIA exams.
Efficiency: Prime rating on 15 of 21 benchmarks in opposition to Gemma 4 E4B, Qwen3.5-4B and Nemotron 3 Nano 4B (NVIDIA-measured).
Greatest: IHEval 84.8, versus 67.0 for Gemma 4 E4B and 60.6 for Qwen3.5-4B.
Worst: Tau2-Bench 55.1, about 20 factors behind Qwen3.5-4B at 75.1.

What’s Nemotron 3.5 Bolt E4B?

Nemotron 3.5 Bolt E4B is a small language mannequin from NVIDIA that powers native conversational brokers. It combines pure dialogue with multi-turn instrument calling. NVIDIA lists 3 goal workloads: gadget and voice assistants, interactive characters, and embedded or industrial brokers.

The analysis report reveals a Reachy Mini robotic working on a Jetson Orin Nano 8GB. In that demo, Cosmos 3 Edge handles imaginative and prescient whereas Bolt handles speech replies and gear calls. NVIDIA crew additionally calls Bolt the third-generation small mannequin for NVIDIA ACE recreation characters.

Core options from the launch publish:

Multi-turn instrument use, with steerage for asynchronous instrument calls.
A conversational_mode flag for brief, speech-ready replies.
Configurable reasoning, plus a runtime considering price range to cap latency.
64K context for historical past, instrument definitions and software state.

How was Bolt E4B educated?

NVIDIA skipped supervised fine-tuning solely. As per the analysis report, SFT improved instrument calling however damage Gemma’s conversational high quality and verbosity. As an alternative, the analysis crew used RLVR and RLHF by means of NeMo-RL and NeMo-Health club.

Each coaching pattern ran in each reasoning and non-reasoning modes. Every batch stored a 2:1 ratio of non-reasoning to reasoning samples. Size penalties within the GenRM surroundings stopped solutions from rising throughout coaching. The datasets and the full recipe are public.

For deployment, NVIDIA quantized the BF16 checkpoint to Q4_K_M with importance-matrix calibration. Calibration used about 1M tokens of RL rollouts. The discharge additionally ships a DFlash drafter for speculative decoding. On an RTX 5070, Bolt with DFlash reached 272.6 tokens/s. Gemma 4 E4B with its MTP drafter reached 225.9 tokens/s, so Bolt is 1.21x quicker.

How does Nemotron 3.5 Bolt E4B carry out on benchmarks?

NVIDIA grouped 21 benchmarks into 5 functionality areas. All scores listed here are BF16, measured by NVIDIA in its personal NeMo-Health club harness. Bolt led 15 of 21 rows in reasoning mode.

NVIDIA additionally reviews roughly 40% increased common scores than Nemotron 3 Nano 4B. It claims as much as 80% fewer output tokens in reasoning-enabled comparisons. Individually, Scale Labs ranked Bolt first amongst small fashions on MultiChallenge at 56.03%, per an X publish.

<div class="pane on" data-p="0">
<div class="tiles">
<div class="tile"><div class="television">8B / 4B</div><div class="tl">whole / efficient parameters</div></div>
<div class="tile"><div class="television">64K</div><div class="tl">context window (tokens)</div></div>
<div class="tile"><div class="television">~4GB</div><div class="tl">goal VRAM envelope</div></div>
<div class="tile"><div class="television">15/21</div><div class="tl">benchmarks led (NVIDIA-measured)</div></div>
</div>
<p class="def"><b>Nemotron 3.5 Bolt E4B</b> is a small language mannequin from NVIDIA for native conversational brokers. It mixes pure dialogue with multi-turn instrument calling, and runs on GeForce RTX, Jetson, RTX Spark and DGX Spark.</p>
<div class="reveal">
<button class="rv"><div class="q">Why "4GB"? ▸</div><div class="a">It’s the reminiscence goal. NVIDIA measured 3.1 to 4.3 GB VRAM on RTX. The Q4_K_M GGUF file is 6.18 GB.</div></button>
<button class="rv"><div class="q">What’s new vs Gemma? ▸</div><div class="a">RL-only post-training for instrument use, instruction hierarchy and concise replies. No SFT stage.</div></button>
<button class="rv"><div class="q">Who’s it for? ▸</div><div class="a">Voice and gadget assistants, recreation NPCs, and robotic or industrial interfaces.</div></button>
</div>
</div>

<div class="pane" data-p="1">
<div class="circulation">
<div class="node on">Gemma 4 E4B</div><div class="arr on"></div>
<div class="node">RLVR + RLHF</div><div class="arr"></div>
<div class="node">Verbosity management</div><div class="arr"></div>
<div class="node">Q4_K_M imatrix</div><div class="arr"></div>
<div class="node">DFlash drafter</div>
</div>
<div class="stepbox">
<div class="stepn" id="mtp-sn">Step 1 of 5</div>
<div class="stept" id="mtp-st"></div>
<p class="stepd" id="mtp-sd"></p>
</div>
<div class="ctl">
<button class="btn ghost" id="mtp-prev">Again</button>
<div class="dots"><span class="dot on"></span><span class="dot"></span><span class="dot"></span><span class="dot"></span><span class="dot"></span></div>
<button class="btn" id="mtp-next">Subsequent</button>
</div>
</div>

<div class="pane" data-p="2">
<div class="chips" id="mtp-chips"></div>
<div id="mtp-bars"></div>
<div class="word">BF16 scores measured by NVIDIA in its NeMo-Health club harness. Vendor-reported scores might differ. Inexperienced worth marks the very best in every row.</div>
</div>

<div class="pane" data-p="3">
<div class="devs" id="mtp-devs"></div>
<div class="meters">
<div class="meter"><div class="mv" id="mtp-vram">0</div><div class="mu">VRAM (GB)</div><div class="mbar"><div class="mfill" id="mtp-vb"></div></div></div>
<div class="meter"><div class="mv" id="mtp-ttft">0</div><div class="mu">Time to first token (ms)</div><div class="mbar"><div class="mfill" id="mtp-tb"></div></div></div>
<div class="meter"><div class="mv" id="mtp-tps">0</div><div class="mu">Technology (tokens/s)</div><div class="mbar"><div class="mfill" id="mtp-pb"></div></div></div>
</div>
<div class="verdict" id="mtp-verdict"></div>
<div class="word">Q4_K_M, llama.cpp (CUDA), batch measurement 1, 3K enter / 50 output tokens, no cached context. Supply: NVIDIA.</div>
</div>

<div class="pane" data-p="4">
<div class="two">
<div class="col w"><h4>The place it wins</h4><ul>
<li>IHEval 84.8 vs 67.0 (Gemma 4 E4B) and 60.6 (Qwen3.5-4B)</li>
<li>MultiChallenge 53.1, prime of the 4 fashions</li>
<li>RelyToolBench 63.2 vs 43.3 for Qwen3.5-4B</li>
<li>1.21x Gemma’s MTP throughput with DFlash on RTX 5070</li>
</ul></div>
<div class="col l"><h4>The place it trails</h4><ul>
<li>Tau2-Bench 55.1 vs 75.1 for Qwen3.5-4B</li>
<li>RULER 64K 87.0 vs 95.1 for Qwen3.5-4B</li>
<li>MMMU-Professional Imaginative and prescient 42.0 vs 47.6 for Gemma 4 E4B</li>
<li>5,785 ms first token on Jetson Orin Nano at 3K enter</li>
</ul></div>
</div>
</div>

<div class="ft"><span>Sources: <a href="https://analysis.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB/" goal="_blank" rel="noopener">NVIDIA weblog</a> · <a href="https://analysis.nvidia.com/labs/nemotron/recordsdata/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-Technical-Report.pdf" goal="_blank" rel="noopener">Tech report</a> · <a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16" goal="_blank" rel="noopener">Mannequin card</a></span><span class="model">© Marktechpost</span></div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-bolt-e4b’);
perform publish(){attempt{var h=R.offsetHeight+40;guardian.postMessage({mtpEmbed:’bolt-e4b’,peak:h},’*’);}catch(e){}}
var tabs=R.querySelectorAll(‘.tab’),panes=R.querySelectorAll(‘.pane’);
tabs.forEach(perform(t){t.addEventListener(‘click on’,perform(){var i=+t.dataset.t;tabs.forEach(perform(x){x.classList.toggle(‘on’,x===t)});panes.forEach(perform(p){p.classList.toggle(‘on’,+p.dataset.p===i)});if(i===2)drawBars();if(i===3)showDev(curDev);setTimeout(publish,60);});});
R.querySelectorAll(‘.rv’).forEach(perform(b){b.addEventListener(‘click on’,perform(){b.classList.toggle(‘open’);setTimeout(publish,380);});});
var steps=[
[‘Start from Gemma 4 E4B’,’Bolt begins as Google’s Gemma-4-E4B-it: 8B total parameters, 4B effective. NVIDIA keeps the backbone and changes behavior.’],
[‘RL-only post-training’,’RLVR and RLHF run in NeMo-RL and NeMo-Gym. Environments cover tool calling, instruction following, instruction hierarchy and structured outputs. SFT was dropped because it hurt conversational quality.’],
[‘Train both modes, control length’,’Every sample runs with reasoning on and off, at a 2:1 non-reasoning to reasoning ratio. Group-relative length penalties keep answers short. NVIDIA reports up to 80% fewer output tokens than peers.’],
[‘Quantize to Q4_K_M’,’The BF16 checkpoint is quantized with llama.cpp using importance-matrix calibration on about 1M tokens of RL rollouts. This build targets a ~4GB VRAM envelope.’],
[‘Speed up with DFlash’,’A parallel drafter proposes 7 tokens in one pass; Bolt verifies them. On RTX 5070 it reaches 272.6 tokens/s vs 225.9 for Gemma with MTP, a 1.21x gain.’]];
var si=0,nodes=R.querySelectorAll(‘.node’),arrs=R.querySelectorAll(‘.arr’),dots=R.querySelectorAll(‘.dot’);
perform step(){doc.getElementById(‘mtp-sn’).textContent=”Step “+(si+1)+’ of 5′;doc.getElementById(‘mtp-st’).textContent=steps[si][0];doc.getElementById(‘mtp-sd’).textContent=steps[si][1];
nodes.forEach(perform(n,i){n.classList.toggle(‘on’,i<=si)});arrs.forEach(perform(a,i){a.classList.toggle(‘on’,i<si)});dots.forEach(perform(d,i){d.classList.toggle(‘on’,i===si)});
doc.getElementById(‘mtp-prev’).disabled=si===0;doc.getElementById(‘mtp-next’).textContent=si===4?’Restart’:’Subsequent’;setTimeout(publish,60);}
doc.getElementById(‘mtp-prev’).addEventListener(‘click on’,perform(){if(si>0){si–;step();}});
doc.getElementById(‘mtp-next’).addEventListener(‘click on’,perform(){si=si===4?0:si+1;step();});
step();
var M=[‘Bolt E4B’,’Nemotron 3 Nano 4B’,’Qwen3.5-4B’,’Gemma 4 E4B’];
var B={‘Dialog’:[[‘MultiChallenge’,’consistency across turns’,[53.1,35.3,51.8,42.1]],[‘RoleMRC’,’role identity adherence’,[80.6,47.0,65.8,74.6]],[‘EQ-Bench v3′,’emotional intelligence’,[39.4,19.3,32.9,41.1]]],
‘Device use’:[[‘BFCL-v3′,’correct tools and arguments’,[68.5,61.1,70.9,66.4]],[‘Tau2-Bench’,’multi-turn customer tasks’,[55.1,29.2,75.1,38.5]],[‘When2Call-MT’,’call, clarify or abstain’,[63.4,27.7,37.8,62.9]]],
‘Instruction’:[[‘IHEval’,’instruction hierarchy’,[84.8,51.3,60.6,67.0]],[‘IFBench’,’precise output constraints’,[63.8,60.1,58.9,42.8]],[‘StructEval-T’,’valid structured outputs’,[80.4,63.5,76.0,79.3]]],
‘Context’:[[‘RULER 64K’,’retrieval and reasoning’,[87.0,91.5,95.1,86.5]],[‘LongBench v2 (48K)’,’long prompts’,[50.5,31.4,59.9,49.3]],[‘LongMemEval’,’recall across conversations’,[72.1,25.2,66.8,70.9]]],
‘Grounding’:[[‘GaRAGe’,’facts from supplied sources’,[42.6,25.2,24.7,34.7]],[‘RAGTruth’,’avoiding unsupported claims’,[77.6,71.4,75.6,77.0]],[‘RelyToolBench’,’tool grounding’,[63.2,41.2,43.3,61.6]]]};
var cat=”Dialog”,chips=doc.getElementById(‘mtp-chips’);
Object.keys(B).forEach(perform(okay){var c=doc.createElement(‘button’);c.className=”chip”+(okay===cat?’ on’:”);c.textContent=okay;c.addEventListener(‘click on’,perform(){cat=okay;chips.querySelectorAll(‘.chip’).forEach(perform(x){x.classList.toggle(‘on’,x===c)});drawBars();});chips.appendChild(c);});
perform drawBars(){var field=doc.getElementById(‘mtp-bars’);field.innerHTML=”;B[cat].forEach(perform(b){var mx=Math.max.apply(null,b[2]);var d=doc.createElement(‘div’);d.className=”bench”;var h=”<div class="bname">”+b[0]+’ <span>’+b[1]+'</span></div>’;
b[2].forEach(perform(v,i){h+='<div class="row’+(i===0?’ bolt’:”)+(v===mx?’ prime’:”)+’"><div class="m">’+M[i]+'</div><div class="monitor"><div class="fill" data-w="’+v+’"></div></div><div class="v">’+v.toFixed(1)+'</div></div>’;});
d.innerHTML=h;field.appendChild(d);});requestAnimationFrame(perform(){requestAnimationFrame(perform(){field.querySelectorAll(‘.fill’).forEach(perform(f){f.model.width=f.dataset.w+’%’;});});});setTimeout(publish,60);}
var D=[[‘RTX 5070′,’12GB’,3.5,548,133.4,’Fast enough for live voice turns, with VRAM left for speech models.’],[‘RTX Spark’,’32GB’,3.1,930,67.0,’Lowest VRAM use in NVIDIA’s tests, under 1 second to first token.’],[‘Jetson AGX Orin’,’32GB’,4.2,3687,24.0,’Workable for robots and kiosks; expect multi-second first replies at 3K input.’],[‘Jetson Orin Nano’,’8GB’,4.2,5785,15.5,’Fits the 8GB board, but first-token latency is the bottleneck at 3K input.’]];
var curDev=0,devs=doc.getElementById(‘mtp-devs’);
D.forEach(perform(d,i){var b=doc.createElement(‘button’);b.className=”dev”+(i===0?’ on’:”);b.innerHTML=d[0]+'<small>’+d[1]+'</small>’;b.addEventListener(‘click on’,perform(){curDev=i;devs.querySelectorAll(‘.dev’).forEach(perform(x){x.classList.toggle(‘on’,x===b)});showDev(i);});devs.appendChild(b);});
perform rely(el,to,dec){var s=efficiency.now(),dur=700;perform f(t){var p=Math.min(1,(t-s)/dur);p=1-Math.pow(1-p,3);var v=to*p;el.textContent=dec?v.toFixed(dec):Math.spherical(v).toLocaleString(‘en-US’);if(p<1)requestAnimationFrame(f);}requestAnimationFrame(f);}
perform showDev(i){var d=D[i];rely(doc.getElementById(‘mtp-vram’),d[2],1);rely(doc.getElementById(‘mtp-ttft’),d[3],0);rely(doc.getElementById(‘mtp-tps’),d[4],1);
doc.getElementById(‘mtp-vb’).model.width=(d[2]/6*100)+’%’;doc.getElementById(‘mtp-tb’).model.width=(d[3]/6000*100)+’%’;doc.getElementById(‘mtp-pb’).model.width=(d[4]/140*100)+’%’;doc.getElementById(‘mtp-verdict’).textContent=d[5];setTimeout(publish,60);}
window.addEventListener(‘load’,publish);window.addEventListener(‘resize’,publish);setTimeout(publish,300);
})();
</script>
“>
(perform(){window.addEventListener(“message”,perform(e){var d=e.knowledge;if(d&&d.mtpEmbed===”bolt-e4b”&&d.peak){var f=doc.getElementById(“mtp-bolt-e4b-frame”);if(f&&e.supply===f.contentWindow){f.model.peak=d.peak+”px”;}}});})();

Bolt doesn’t win in all places. Qwen3.5-4B leads on Tau2-Bench (75.1 vs 55.1) and RULER 64K (95.1 vs 87.0). It additionally leads LongBench v2 (59.9 vs 50.5) and BFCL-v3 (70.9 vs 68.5). Gemma 4 E4B retains a slim edge on EQ-Bench v3 (41.1 vs 39.4).

How does Bolt E4B examine with different small fashions?

Characteristic	Nemotron 3.5 Bolt E4B	Gemma 4 E4B	Qwen3.5-4B	Nemotron 3 Nano 4B
Parameters	8B whole, 4B efficient	8B with embeddings, 4.5B efficient	4B (language mannequin)	3.97B
Structure	Gemma 4 E4B, RL post-trained	Dense with Per-Layer Embeddings	Gated DeltaNet + consideration hybrid	Mamba-2 + Transformer hybrid
Context	64K	128K	262K	262K
License	OpenMDW-1.1	Apache 2.0	Apache 2.0	NVIDIA Nemotron Open Mannequin License
Inputs	Textual content (card); NIM additionally accepts picture and audio	Textual content, picture, audio	Textual content, picture, video	Textual content
IHEval*	84.8	67.0	60.6	51.3
MultiChallenge*	53.1	42.1	51.8	35.3
Tau2-Bench*	55.1	38.5	75.1	29.2
RULER 64K*	87.0	86.5	95.1	91.5
Availability	Open weights; free API on construct.nvidia.com	Open weights	Open weights	Open weights

*Benchmark rows are NVIDIA-measured BF16 scores from the Bolt launch publish. Vendor-reported scores differ.

What {hardware} does it run on?

NVIDIA measured the Q4_K_M construct with llama.cpp at batch measurement 1. The figures under use 3K enter and 50 output tokens.

Gadget	VRAM (GB)	Time to first token (ms)	Tokens/s
RTX 5070 (12GB)	3.5	548	133.4
RTX Spark (32GB)	3.1	930	67.0
Jetson AGX Orin (32GB)	4.2	3,687	24.0
Jetson Orin Nano (8GB)	4.2	5,785	15.5

The ‘4GB’ label describes a goal reminiscence envelope, not the file measurement. The Q4_K_M GGUF on Hugging Face is 6.18 GB.

Bolt is obtainable on Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab. It additionally ships as an NVIDIA NIM and thru the In-Sport Inferencing SDK.