Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Mannequin

Nanbeige LLM Lab has launched Nanbeige4.3-33B on Hugging Face underneath the Apache 2.0 license. Nanbeige4.3-33B is a 33B dense language mannequin constructed for agentic coding and office-style ‘cowork’ duties. The lab sits contained in the recruitment platform BOSS Zhipin. It additionally revealed Nanbeige4.3-33B-Base for customized post-training. The mannequin helps a 524,288-token context window. Within the lab’s personal desk, it leads 17 of 20 benchmarks in opposition to 3 open fashions close to 30B.

TL;DR

Dimension: 33B parameters, dense (no MoE), 524,288-token (512K) context, BF16 weights.
Runs on: Multi-GPU servers. The cardboard’s SGLang and vLLM recipes use 4-way tensor parallelism. BF16 weights want about 66 GB (our arithmetic). 1 neighborhood quantization is listed.
Efficiency: Leads 17 of 20 rows in Nanbeige’s personal comparability in opposition to Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B.
Greatest: 80.9 on Terminal-Bench 2.1, versus 73.0 for Qwen3.8-27B.
Worst: 23.8 on Terminal-Bench 4.0 (its lowest rating). It additionally trails Qwen3.8-27B on SWE-Bench Professional, 60.3 vs 61.7.

What’s Nanbeige4.3-33B?

Nanbeige4.3-33B is the flagship ‘trainer’ mannequin of the Nanbeige4.3 sequence. The lab’s said technique is to construct a powerful trainer first. It then plans to distill agentic expertise into smaller fashions, beginning with Nanbeige4.3-3B.

How is Nanbeige4.3-33B constructed?

The config.json declares a regular LlamaForCausalLM structure. There isn’t any professional routing, so all 33B parameters run on each token.

52 layers, hidden dimension 7,168
56 question heads and eight key-value heads (grouped-query consideration), head dimension 128
FFN intermediate dimension 22,016, vocabulary of 166,912 tokens
RoPE theta of 120,000,000 and 524,288 max positions

The chat template exposes 3 controls. enable_thinking toggles reasoning and is on by default. preserve_thinking retains reasoning from earlier turns. Nanbeige recommends it for multi-turn instrument use and code brokers. Passing instruments allows instrument calling, and an XML name format is really helpful.

How does Nanbeige4.3-33B carry out on benchmarks?

Nanbeige in contrast the mannequin with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Each mannequin ran at its highest considering effort, with shared hyperparameters and harnesses. All scores on this part are self-reported.

Agentic coding: It scores 80.9 on Terminal-Bench 2.1, in opposition to 73.0 for Qwen3.8-27B. DeepSWE reveals the widest hole: 65.5 versus 42.5. On Terminal-Bench 4.0, it scores 23.8 whereas Gemma4-31B scores 0.0.
Agentic cowork: It posts 54.2 on Automation Bench versus 38.3, and 44.8 on JobBench versus 34.5. On YC-bench, simulated earnings attain $1,025,437 versus $641,290.
Reasoning: It experiences 95.2 on AIME 2026 and 84.0 on AA-LCR.

It loses 3 of 20 rows. Qwen3.8-27B leads SWE-Bench Professional (61.7 vs 60.3) and GDPval (86.4 vs 84.8). Gemma4-31B leads GPQA Diamond in Nanbeige’s desk, 91.3 vs 89.5.

The Base mannequin beats Gemma4-31B-Base on all 8 reported duties. Examples embody MMLU-Professional (66.9 vs 59.3) and GPQA (62.0 vs 41.5).

<div class="tabs" function="tablist">
<button class="tab on" data-p="p1">The way it works</button>
<button class="tab" data-p="p2">Benchmarks</button>
<button class="tab" data-p="p3">Reminiscence math</button>
<button class="tab" data-p="p4">Caveats</button>
</div>

<div class="pane on" id="p1">
<div class="circulate">
<div class="node on" data-s="0"><div class="n">STEP 1</div><div class="t">Base mannequin</div><div class="d">Nanbeige4.3-33B-Base, launched for customized post-training</div></div>
<div class="node" data-s="1"><div class="n">STEP 2</div><div class="t">Trainer agent</div><div class="d">Submit-trained Nanbeige4.3-33B with considering and instruments</div></div>
<div class="node" data-s="2"><div class="n">STEP 3</div><div class="t">Agent harness</div><div class="d">Runs inside harnesses like Terminus 2, mini-swe-agent, Claude Code</div></div>
<div class="node" data-s="3"><div class="n">STEP 4</div><div class="t">Distillation</div><div class="d">Deliberate: agentic expertise transfer into Nanbeige4.3-3B</div></div>
</div>
<div class="rail"><i id="nbgRail"></i></div>
<div class="element" id="nbgDetail"></div>
<div class="ctl"><button class="btn ghost" id="nbgPrev"> Again</button><button class="btn" id="nbgNext">Subsequent step </button><button class="btn ghost" id="nbgPlay">Auto-play</button></div>
<div class="spec">
<div><b>52</b><span>layers</span></div>
<div><b>56 / 8</b><span>question / KV heads (GQA)</span></div>
<div><b>7,168</b><span>hidden dimension</span></div>
<div><b>524,288</b><span>max positions</span></div>
</div>
</div>

<div class="pane" id="p2">
<div class="seg" id="nbgSeg">
<button class="on" data-c="coding">Agentic coding</button>
<button data-c="cowork">Agentic cowork</button>
<button data-c="cause">Reasoning</button>
<button data-c="base">Base mannequin</button>
</div>
<div class="legend" id="nbgLegend"></div>
<div id="nbgBars"></div>
<p class="word">All scores self-reported by Nanbeige on its Hugging Face mannequin playing cards, at every mannequin’s highest considering effort with shared harnesses. Not independently verified.</p>
</div>

<div class="pane" id="p3">
<div class="calc">
<div>
<div class="discipline"><label>Context size <b id="nbgCtxL">131,072 tokens</b></label><enter kind="vary" id="nbgCtx" min="0" max="6" step="1" worth="4"></div>
<div class="discipline"><label>Concurrent sequences <b id="nbgSeqL">1</b></label><enter kind="vary" id="nbgSeq" min="1" max="8" step="1" worth="1"></div>
<div class="seg" id="nbgW"><button class="on" data-w="2">Weights BF16</button><button data-w="1">Weights 8-bit</button></div>
<div class="seg" id="nbgK"><button class="on" data-k="2">KV cache BF16</button><button data-k="1">KV cache FP8</button></div>
<p class="word">Arithmetic from config.json: 52 layers × 8 KV heads × 128 dims × 2 (Ok and V) × bytes per worth. That’s 212,992 bytes (about 208 KiB) per token in BF16. Weights use 33B parameters. Excludes activations and framework overhead.</p>
</div>
<div class="stack">
<div class="kv"><span>Estimated GPU reminiscence</span></div>
<div class="whole" id="nbgTot">0 GB</div>
<div class="stackbar"><i id="nbgBw" model="background:#5fd3bd"></i><i id="nbgBk" model="background:#7c8cf8"></i></div>
<div class="kv"><span><i model="show:inline-block;width:9px;top:9px;background:#5fd3bd;border-radius:2px"></i> Weights</span><b id="nbgWv"></b></div>
<div class="kv"><span><i model="show:inline-block;width:9px;top:9px;background:#7c8cf8;border-radius:2px"></i> KV cache</span><b id="nbgKv"></b></div>
<div class="kv"><span>Per 80 GB GPU (perfect match)</span><b id="nbgGpu"></b></div>
<p class="word">The mannequin card’s SGLang and vLLM recipes use 4-way tensor parallelism.</p>
</div>
</div>
</div>

<div class="pane" id="p4">
<desk>
<tr><th>Row in Nanbeige desk</th><th>Nanbeige lists</th><th>Official card</th><th></th></tr>
<tr><td>Muse-Glimmer-30B, SWE-Bench Professional</td><td class="unhealthy">36.9</td><td>51.2 (Meta)</td><td><span class="flag">appears swapped</span></td></tr>
<tr><td>Gemma4-31B, SWE-Bench Professional</td><td class="unhealthy">51.2</td><td>36.9 (Meta-reported)</td><td><span class="flag">appears swapped</span></td></tr>
<tr><td>Gemma4-31B, GPQA Diamond</td><td class="unhealthy">91.3</td><td>84.3 (Google)</td><td><span class="flag">battle</span></td></tr>
<tr><td>Muse-Glimmer-30B, Terminal-Bench 2.1</td><td class="unhealthy">61.8</td><td>51.7 (Meta)</td><td><span class="flag">harness?</span></td></tr>
<tr><td>Qwen3.8-27B, GPQA Diamond</td><td>89.2</td><td>89.2 (Qwen)</td><td>match</td></tr>
<tr><td>Muse-Glimmer-30B, AIME 2026</td><td>94.7</td><td>94.7 (Meta)</td><td>match</td></tr>
</desk>
<p class="word">Nanbeige re-ran each mannequin with its personal harness, which explains some gaps. Different limits: text-only enter, dense compute on each token, and Nanbeige4.3-3B isn’t but launched.</p>
</div>

<div class="ft"><span>Sources: <a href="https://huggingface.co/Nanbeige/Nanbeige4.3-33B" goal="_blank" rel="noopener">Nanbeige mannequin card</a> · <a href="https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base" goal="_blank" rel="noopener">Base card</a> · competitor HF playing cards</span><span class="model">© Marktechpost</span></div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-nbg43′);
perform resize(){strive{guardian.postMessage({kind:’mtp-nbg43-resize’,top:R.offsetHeight+40},’*’)}catch(e){}}
// tabs
R.querySelectorAll(‘.tab’).forEach(perform(t){t.addEventListener(‘click on’,perform(){
R.querySelectorAll(‘.tab’).forEach(perform(x){x.classList.take away(‘on’)});t.classList.add(‘on’);
R.querySelectorAll(‘.pane’).forEach(perform(p){p.classList.take away(‘on’)});
doc.getElementById(t.dataset.p).classList.add(‘on’);
if(t.dataset.p===’p2′)drawBars(cur);
setTimeout(resize,50);setTimeout(resize,400);
})});
// circulate
var steps=[
‘<b>Base model.</b> Nanbeige4.3-33B-Base is a dense LlamaForCausalLM checkpoint. Nanbeige reports it beats Gemma4-31B-Base on all 8 listed tasks, for example MMLU-Pro 66.9 vs 59.3.’,
‘<b>Teacher agent.</b> The post-trained model thinks by default (<code>enable_thinking</code>). <code>preserve_thinking</code> keeps earlier reasoning for multi-turn tool use. Nanbeige recommends the XML tool-call format.’,
‘<b>Agent harness.</b> Nanbeige evaluated it in Terminus 2, mini-swe-agent v2.4.6 and the Claude Code harness. Recommended agent sampling: temperature 1.0, top_p 0.95, 65,536 max new tokens.’,
‘<b>Distillation.</b> The 33B model is the teacher for the Nanbeige4.3 series. The first planned student is Nanbeige4.3-3B. No release date is given.’
];
var s=0,timer=null;
perform setStep(i){s=(i+4)%4;R.querySelectorAll(‘.node’).forEach(perform(n,okay){n.classList.toggle(‘on’,okay<=s)});
doc.getElementById(‘nbgRail’).model.width=((s+1)*25)+’%’;doc.getElementById(‘nbgDetail’).innerHTML=steps[s];resize()}
R.querySelectorAll(‘.node’).forEach(perform(n){n.addEventListener(‘click on’,perform(){cease();setStep(+n.dataset.s)})});
perform cease(){if(timer){clearInterval(timer);timer=null;doc.getElementById(‘nbgPlay’).textContent=”Auto-play”}}
doc.getElementById(‘nbgNext’).onclick=perform(){cease();setStep(s+1)};
doc.getElementById(‘nbgPrev’).onclick=perform(){cease();setStep(s-1)};
doc.getElementById(‘nbgPlay’).onclick=perform(){if(timer){cease();return}this.textContent=”Pause”;timer=setInterval(perform(){setStep(s+1)},2200)};
setStep(0);
// bars
var M={n:[‘Nanbeige4.3-33B’,’#5fd3bd’],q:[‘Qwen3.8-27B’,’#7c8cf8′],m:[‘Muse-Glimmer-30B’,’#f0a35e’],g:[‘Gemma4-31B’,’#6aa9f4′]};
var D={
coding:[[‘Terminal-Bench 2.1’,80.9,73.0,61.8,47.2],[‘DeepSWE’,65.5,42.5,4.4,2.7],[‘SWE-Bench Pro’,60.3,61.7,36.9,51.2],[‘SkillsBench’,61.1,56.4,39.7,28.9],[‘Terminal-Bench 4.0’,23.8,6.3,1.6,0.0]],
cowork:[[‘Automation Bench’,54.2,38.3,15.2,9.7],[‘JobBench’,44.8,34.5,16.6,9.0],[‘GDPval’,84.8,86.4,75.8,71.4],[‘Toolathlon-Verified’,70.4,68.5,47.2,27.8],[‘DeepPlanning’,63.3,49.2,21.4,19.4]],
cause:[[‘AIME 2026’,95.2,94.5,94.7,89.2],[‘GPQA Diamond’,89.5,89.2,83.5,91.3],[‘AA-LCR’,84.0,82.0,83.3,72.0]],
base:[[‘MMLU-Pro’,66.9,null,null,59.3],[‘GPQA’,62.0,null,null,41.5],[‘GSM8K’,92.9,null,null,88.9],[‘Math’,57.3,null,null,43.6],[‘CMMLU’,86.8,null,null,76.9]]
};
var cur=”coding”;
perform drawBars(c){cur=c;var keys=c===’base’?[‘n’,’g’]:[‘n’,’q’,’m’,’g’];
var names=c===’base’?{n:’Nanbeige4.3-33B-Base’,g:’Gemma4-31B-Base’}:{};
doc.getElementById(‘nbgLegend’).innerHTML=keys.map(perform(okay)M[k][0])+'</span>’).be a part of(”);
var idx={n:1,q:2,m:3,g:4},h=””;
D[c].forEach(perform(r){h+='<div class="bench"><div class="bn">’+r[0]+'</div>’;
keys.forEach(perform(okay));
h+='</div>’});
var field=doc.getElementById(‘nbgBars’);field.innerHTML=h;
requestAnimationFrame(perform(){requestAnimationFrame(perform(){field.querySelectorAll(‘.fill’).forEach(perform(f){f.model.width=f.dataset.w+’%’})})});
resize();setTimeout(resize,300)}
R.querySelectorAll(‘#nbgSeg button’).forEach(perform(b){b.onclick=perform(){R.querySelectorAll(‘#nbgSeg button’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawBars(b.dataset.c)}});
drawBars(‘coding’);
// calc
var CT=[8192,16384,32768,65536,131072,262144,524288],wb=2,kb=2;
perform fmt(n){return n.toLocaleString(‘en-US’)}
perform calc(){var ctx=CT[+document.getElementById(‘nbgCtx’).value],seq=+doc.getElementById(‘nbgSeq’).worth;
doc.getElementById(‘nbgCtxL’).textContent=fmt(ctx)+’ tokens’;doc.getElementById(‘nbgSeqL’).textContent=seq;
var w=33e9*wb/1e9,kv=52*8*128*2*kb*ctx*seq/1e9,t=w+kv;
doc.getElementById(‘nbgTot’).textContent=t.toFixed(1)+’ GB’;
doc.getElementById(‘nbgWv’).textContent=w.toFixed(1)+’ GB’;doc.getElementById(‘nbgKv’).textContent=kv.toFixed(1)+’ GB’;
doc.getElementById(‘nbgGpu’).textContent=Math.ceil(t/80)+’ × 80 GB’;
doc.getElementById(‘nbgBw’).model.width=(w/t*100)+’%’;doc.getElementById(‘nbgBk’).model.width=(kv/t*100)+’%’}
[‘nbgCtx’,’nbgSeq’].forEach(perform(id){doc.getElementById(id).addEventListener(‘enter’,calc)});
R.querySelectorAll(‘#nbgW button’).forEach(perform(b){b.onclick=perform(){R.querySelectorAll(‘#nbgW button’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);wb=+b.dataset.w;calc()}});
R.querySelectorAll(‘#nbgK button’).forEach(perform(b){b.onclick=perform(){R.querySelectorAll(‘#nbgK button’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);kb=+b.dataset.okay;calc()}});
calc();
window.addEventListener(‘load’,resize);window.addEventListener(‘resize’,resize);setTimeout(resize,200);
})();
</script>
</physique></html>
” model=”width:100%!necessary;top:600px;border:0!necessary;background:clear!necessary;show:block;” loading=”lazy”>

How does it evaluate with rivals’ personal mannequin playing cards?

Spec	Nanbeige4.3-33B	Qwen3.8-27B	Muse-Glimmer-30B	Gemma 4 31B
Developer	Nanbeige LLM Lab (BOSS Zhipin)	Qwen (Alibaba)	Meta Superintelligence Lab	Google DeepMind
Parameters	33B dense	27B dense LM plus imaginative and prescient encoder	~29.6B dense, incl. ~1.8B imaginative and prescient encoder	30.7B dense
Context	524,288	262,144 native, as much as 1M	131,072+	256K
Enter modality	Textual content	Textual content, picture, video	Textual content, picture	Textual content, picture
License	Apache 2.0	Apache 2.0	Apache 2.0	Apache 2.0
Terminal-Bench 2.1	80.9*	73.0	51.7	Not disclosed
SWE-Bench Professional	60.3*	61.7	51.2	Not disclosed
GPQA Diamond	89.5*	89.2	83.5 (AA)	84.3
AIME 2026	95.2*	Not disclosed	94.7	89.2
Native {hardware} steerage	4-GPU tensor-parallel recipe	Not disclosed	24 GB or 32 GB VRAM at ~4-bit	Shopper GPUs and workstations

*Self-reported by Nanbeige. Competitor values come from every lab’s personal Hugging Face mannequin card.

Some competitor figures in Nanbeige’s desk differ from what these labs publish:

Meta experiences Muse-Glimmer-30B at 51.2 on SWE-Bench Professional and Gemma4-31B at 36.9. Nanbeige lists the reverse.
Google lists Gemma 4 31B at 84.3 on GPQA Diamond. Nanbeige’s desk reveals 91.3.
Meta experiences Muse-Glimmer-30B at 51.7 on Terminal-Bench 2.1. Nanbeige measured 61.8.

Totally different harnesses clarify some gaps. The SWE-Bench Professional pair appears swapped. Deal with these rows as provisional till unbiased runs seem.

How do you deploy Nanbeige4.3-33B?

The mannequin card gives Transformers, SGLang and vLLM recipes. Its server examples use 4-way tensor parallelism with Qwen3 reasoning and tool-call parsers. Really helpful agent settings are temperature 1.0, top_p 0.95 and 65,536 max new tokens. Reasoning and chat use temperature 0.6 and 262,144 max new tokens.

By arithmetic from config.json, BF16 weights want about 66 GB. The KV cache provides about 208 KiB per token in BF16. A full 524,288-token sequence subsequently wants roughly 112 GB of KV cache. At publication, no Hugging Face Inference Supplier hosts the mannequin.