This week RadixArk launched DaisyRose-27B, its first open-weight mannequin. DaisyRose-27B is an agentic coding mannequin from RadixArk, post-trained from Qwen3.8-27B with reinforcement studying (RL). The mannequin card web page states that it used RL solely, with no SFT or distillation. It ships below Apache 2.0 as a drop-in alternative for the bottom mannequin. The principle consequence: Terminal Bench 4.1 go@1 rises from 5.6 to 12.0. You may examine the deployment choices and specs right here via our GPULeaderboard.com chrome extension.
TL;DR
- Measurement: 27.8B dense parameters (0.46B imaginative and prescient encoder, 0.42B MTP head), 262,144-token context.
- Runs on: 1 GPU. BF16 is about 52 GB, FP8 about 29 GB, NVFP4 about 23 GB (Blackwell solely). Listed targets: H200, B200, GB300, RTX PRO 6000, RTX 5090, DGX Spark. GGUF additionally obtainable.
- Efficiency: Beats its Qwen3.8-27B base on all 7 reported benchmarks.
- Finest: IFBench 82.4, the highest rating among the many 9 fashions RadixArk in contrast.
- Worst: HLE (textual content solely) 34.3, solely +0.4 over the bottom and beneath Inkling 2 Small’s 44.1.
- Backside line: Finest: Terminal Bench 4.1 greater than doubles after 18 days of RL. Worst: it nonetheless trails a lot bigger MoE fashions on exhausting agentic duties.
What’s DaisyRose-27B?
DaisyRose-27B retains the Qwen3.8-27B structure unchanged. The mannequin card lists 64 decoder layers in a 3:1 sample. Three Gated DeltaNet linear-attention layers alternate with 1 full-attention layer. Hidden measurement is 5,120. A 27-layer imaginative and prescient encoder handles pictures. One multi-token-prediction layer helps speculative decoding. Enter is textual content and pictures, output is textual content, and it’s a considering mannequin.
How was DaisyRose-27B skilled?
RadixArk constructed about 1,200 curated duties with Eigent AI, chosen from 1000’s of candidates. Every job targets a selected failure of the bottom mannequin. Every sandbox try permits as much as 1.5 hours, 100+ software calls and 256k+ output tokens. Agent classes attain as much as 1M tokens, per the weblog.
Coaching ran on Miles, RadixArk’s open-source RL framework, in absolutely asynchronous mode. The run used 144 H200 GPUs for 18 days. Of these, 128 dealt with agent inference and 16 dealt with coaching. Whole compute was 63,000 GPU-hours.
RadixArk reviews 37% MFU on its inner backend. An open-source Megatron-based coach reached 16%. Finish-to-end throughput was 16.8k tokens/s versus 8.3k, a 2.0x speedup. RadixArk says this lower GPU time by 57%.
Sandboxes ran on AgentENV. RadixArk reviews about 944,000 sandboxes from mid-September to mid-November. Peak concurrency hit 1,566. Creation success was 99% over 9 weeks. RadixArk places the associated fee about 80% beneath fashionable cloud sandbox suppliers.
How does DaisyRose-27B carry out on benchmarks?
All scores are self-reported by RadixArk. DaisyRose scores are the upper of its BF16 and FP8 runs.
| Benchmark | DaisyRose-27B | Qwen3.8-27B | Change |
|---|---|---|---|
| Terminal Bench 4.1 (go@1) | 12.0 | 5.6 | +6.4 |
| Terminal Bench 2.1 | 75.5 | 73.0 | +2.5 |
| DeepSWE v1.1 | 46.2 | 42.2 | +4.0 |
| SWE Bench Professional v2 | 94.7 | 91.3 | +3.4 |
| HLE (textual content solely) | 34.3 | 33.9 | +0.4 |
| IFBench | 82.4 | 79.5 | +2.9 |
| GPQA Diamond | 90.7 | 89.2 | +1.5 |
Supply: RadixArk weblog and mannequin card.
The biggest leap is Terminal Bench 4.1, which greater than doubles. Coding positive aspects are regular: +4.0 on DeepSWE v1.1 and +3.4 on SWE Bench Professional v2. Data positive aspects are small, at +0.4 on HLE and +1.5 on GPQA Diamond.
In opposition to bigger fashions, outcomes are combined. MiMo V2.6 Flash (309B whole) leads Terminal Bench 4.1 at 16.1. Inkling 2 Small (594B whole) leads DeepSWE at 72.1 and HLE at 44.1. DaisyRose-27B tops RadixArk’s 9-model desk solely on IFBench.
<html lang="en"><head><meta charset="utf-8"><meta title="viewport" content material="width=device-width,initial-scale=1">
<title>DaisyRose-27B Explainer | Marktechpost</title>
<type>
html,physique{margin:0!essential;padding:0!essential;background:clear!essential;}
#mtp-drose{–ink:#0f0f14;–card:#17171f;–line:#2a2a36;–rose:#e8457a;–gold:#f5c542;–txt:#ecebf2;–mut:#9a98ab;–green:#76B900;
max-width:880px;margin:0 auto!essential;background:var(–ink)!essential;colour:var(–txt)!essential;border:1px strong var(–line)!essential;border-radius:14px;font-family:Inter,system-ui,-apple-system,Segoe UI,Roboto,sans-serif;overflow:hidden;box-sizing:border-box}
#mtp-drose *{box-sizing:border-box}
#mtp-drose hr,#mtp-drose p:empty,#mtp-drose del,#mtp-drose s{show:none!essential}
#mtp-drose .hd{padding:20px 22px 14px;border-bottom:1px strong var(–line)!essential;background:linear-gradient(120deg,rgba(232,69,122,.16),rgba(245,197,66,.06) 60%,clear)!essential}
#mtp-drose .kick{font-size:11px;letter-spacing:.14em;text-transform:uppercase;colour:var(–gold)!essential;font-weight:700}
#mtp-drose h3{margin:6px 0 4px!essential;font-size:22px;line-height:1.25;colour:var(–txt)!essential;font-weight:800}
#mtp-drose .sub{font-size:13.5px;colour:var(–mut)!essential;margin:0}
#mtp-drose .tabs{show:flex;hole:4px;padding:10px 14px 0;border-bottom:1px strong var(–line)!essential;overflow-x:auto;scrollbar-width:none}
#mtp-drose .tab{background:clear!essential;border:0!essential;border-bottom:2px strong clear!essential;colour:var(–mut)!essential;font:600 13px Inter,sans-serif;padding:9px 12px;cursor:pointer;white-space:nowrap;transition:colour .2s}
#mtp-drose .tab:hover{colour:var(–txt)!essential}
#mtp-drose .tab.on{colour:var(–txt)!essential;border-bottom-color:var(–rose)!essential}
#mtp-drose .pane{show:none;padding:20px 22px 22px;animation:drIn .35s ease}
#mtp-drose .pane.on{show:block}
@keyframes drIn{from{opacity:0;rework:translateY(6px)}to{opacity:1;rework:none}}
#mtp-drose .tiles{show:grid;grid-template-columns:repeat(4,1fr);hole:10px}
#mtp-drose .tile{background:var(–card)!essential;border:1px strong var(–line)!essential;border-radius:10px;padding:12px;cursor:pointer;transition:rework .2s,border-color .2s}
#mtp-drose .tile:hover{rework:translateY(-2px);border-color:var(–rose)!essential}
#mtp-drose .tile b{show:block;font-size:22px;colour:var(–gold)!essential;font-weight:800}
#mtp-drose .tile span{font-size:12px;colour:var(–mut)!essential}
#mtp-drose .tile em{show:none;font-style:regular;font-size:12px;colour:var(–txt)!essential;margin-top:6px;line-height:1.4}
#mtp-drose .tile.open em{show:block}
#mtp-drose .lbl{font-size:12px;colour:var(–mut)!essential;margin:18px 0 8px;text-transform:uppercase;letter-spacing:.1em;font-weight:700}
#mtp-drose .layers{show:grid;grid-template-columns:repeat(32,1fr);hole:3px}
#mtp-drose .ly{peak:22px;border-radius:3px;background:#2b2b38!essential;transition:background .25s,rework .25s}
#mtp-drose .ly.lin.lit{background:var(–rose)!essential}
#mtp-drose .ly.att.lit{background:var(–gold)!essential;rework:scaleY(1.25)}
#mtp-drose .legend{show:flex;hole:16px;font-size:12px;colour:var(–mut)!essential;margin-top:8px;flex-wrap:wrap}
#mtp-drose .dot{show:inline-block;width:10px;peak:10px;border-radius:2px;margin-right:6px;vertical-align:-1px}
#mtp-drose .circulate{show:grid;grid-template-columns:repeat(4,1fr);hole:10px;place:relative}
#mtp-drose .st{background:var(–card)!essential;border:1px strong var(–line)!essential;border-radius:10px;padding:12px;opacity:.38;transition:opacity .3s,border-color .3s,box-shadow .3s;place:relative}
#mtp-drose .st.on{opacity:1;border-color:var(–rose)!essential;box-shadow:0 0 0 3px rgba(232,69,122,.18)}
#mtp-drose .st .n{font-size:11px;colour:var(–gold)!essential;font-weight:800}
#mtp-drose .st h4{margin:4px 0!essential;font-size:14px;colour:var(–txt)!essential}
#mtp-drose .st p{margin:0!essential;font-size:12px;colour:var(–mut)!essential;line-height:1.45}
#mtp-drose .observe{peak:6px;background:#23232e!essential;border-radius:4px;margin:14px 0 6px;place:relative;overflow:hidden}
#mtp-drose .observe i{place:absolute;high:0;left:0;peak:100%;background:linear-gradient(90deg,var(–rose),var(–gold))!essential;transition:width .4s}
#mtp-drose .pulse{place:absolute;high:-1px;width:8px;peak:8px;border-radius:50%;background:#fff!essential;animation:drRun 1.6s linear infinite}
@keyframes drRun{from{left:-8px}to{left:100%}}
#mtp-drose .element{background:var(–card)!essential;border-left:3px strong var(–gold)!essential;border-radius:6px;padding:12px 14px;font-size:13.5px;line-height:1.55;colour:var(–txt)!essential;margin-top:10px;min-height:66px}
#mtp-drose .ctrl{show:flex;hole:8px;margin-top:12px;flex-wrap:wrap}
#mtp-drose .btn{background:var(–card)!essential;colour:var(–txt)!essential;border:1px strong var(–line)!essential;border-radius:8px;padding:8px 12px;font:600 12.5px Inter,sans-serif;cursor:pointer;transition:border-color .2s,background .2s}
#mtp-drose .btn:hover{border-color:var(–rose)!essential}
#mtp-drose .btn.on{background:var(–rose)!essential;border-color:var(–rose)!essential;colour:#fff!essential}
#mtp-drose .bars{margin-top:12px}
#mtp-drose .br{show:grid;grid-template-columns:150px 1fr 48px;hole:10px;align-items:heart;margin:6px 0;font-size:12.5px}
#mtp-drose .br .nm{colour:var(–mut)!essential;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}
#mtp-drose .br.me .nm{colour:var(–txt)!essential;font-weight:700}
#mtp-drose .bg{peak:16px;background:#1f1f29!essential;border-radius:4px;overflow:hidden}
#mtp-drose .fl{peak:100%;width:0;background:#4a4a5c!essential;border-radius:4px;transition:width .8s cubic-bezier(.2,.8,.2,1)}
#mtp-drose .br.me .fl{background:linear-gradient(90deg,var(–rose),var(–gold))!essential}
#mtp-drose .br.base .fl{background:#7d7b92!essential}
#mtp-drose .vl{text-align:proper;font-weight:700;colour:var(–txt)!essential}
#mtp-drose .notice{font-size:11.5px;colour:var(–mut)!essential;margin-top:10px;line-height:1.5}
#mtp-drose .grid2{show:grid;grid-template-columns:1fr 1fr;hole:12px}
#mtp-drose .massive{font-size:34px;font-weight:800;colour:var(–gold)!essential}
#mtp-drose pre,#mtp-drose code{background:#0a0a0e!essential;colour:#f3c9d7!essential;border:1px strong var(–line)!essential;font-family:ui-monospace,Menlo,Consolas,monospace!essential;font-size:12px!essential}
#mtp-drose pre{padding:12px!essential;border-radius:8px;overflow-x:auto;margin:10px 0 0!essential;white-space:pre}
#mtp-drose .chips{show:flex;hole:6px;flex-wrap:wrap;margin-top:8px}
#mtp-drose .chip{font-size:12px;padding:5px 9px;border-radius:20px;border:1px strong var(–line)!essential;colour:var(–txt)!essential;background:var(–card)!essential}
#mtp-drose .ft{show:flex;justify-content:space-between;hole:10px;flex-wrap:wrap;padding:11px 22px;border-top:1px strong var(–line)!essential;font-size:11.5px;colour:var(–mut)!essential}
#mtp-drose .ft a{colour:#76B900!essential;text-decoration:none;font-weight:700}
@media(max-width:640px){
#mtp-drose h3{font-size:18px}
#mtp-drose .tiles{grid-template-columns:repeat(2,1fr)}
#mtp-drose .circulate{grid-template-columns:1fr 1fr}
#mtp-drose .grid2{grid-template-columns:1fr}
#mtp-drose .br{grid-template-columns:104px 1fr 40px;font-size:11.5px}
#mtp-drose .layers{grid-template-columns:repeat(16,1fr)}
#mtp-drose .pane{padding:16px}
}
</type></head><physique>
<div id="mtp-drose">
<div class="hd">
<div class="kick">Interactive explainer · RadixArk</div>
<h3>DaisyRose-27B: RL-only post-training on Qwen3.8-27B</h3>
<p class="sub">How it’s constructed, the way it was skilled, the way it scores and methods to run it. All figures from RadixArk's weblog and mannequin card.</p>
</div>
<div class="tabs" position="tablist">
<button class="tab on" data-p="ov">Overview</button>
<button class="tab" data-p="tr">Coaching loop</button>
<button class="tab" data-p="bm">Benchmarks</button>
<button class="tab" data-p="ef">Effectivity</button>
<button class="tab" data-p="dp">Deploy</button>
</div>
<div class="pane on" id="dr-ov">
<div class="tiles">
<div class="tile"><b>27.8B</b><span>dense params</span><em>Features a 0.46B imaginative and prescient encoder and a 0.42B multi-token-prediction head.</em></div>
<div class="tile"><b>262,144</b><span>token context</span><em>Native context per the mannequin card. The weblog cites agent classes as much as 1M tokens.</em></div>
<div class="tile"><b>1 GPU</b><span>single-card deploy</span><em>BF16 about 52 GB, FP8 about 29 GB, NVFP4 about 23 GB.</em></div>
<div class="tile"><b>Apache 2.0</b><span>license</span><em>Identical license because the Qwen3.8-27B base. Drop-in alternative.</em></div>
</div>
<div class="lbl">64 decoder layers · 3:1 hybrid sample (click on to animate)</div>
<div class="layers" id="dr-layers"></div>
<div class="legend"><span><i class="dot" type="background:#e8457a"></i>Gated DeltaNet (linear consideration) × 48</span><span><i class="dot" type="background:#f5c542"></i>Full consideration × 16</span><span>Hidden measurement 5,120 · 27-layer imaginative and prescient encoder · 1 MTP layer</span></div>
</div>
<div class="pane" id="dr-tr">
<div class="circulate" id="dr-flow">
<div class="st on"><div class="n">STEP 1</div><h4>Curate duties</h4><p>~1,200 duties from 1000’s of candidates</p></div>
<div class="st"><div class="n">STEP 2</div><h4>Agent rollouts</h4><p>128 H200s in AgentENV sandboxes</p></div>
<div class="st"><div class="n">STEP 3</div><h4>Async RL replace</h4><p>16 H200s prepare with Miles</p></div>
<div class="st"><div class="n">STEP 4</div><h4>Ship weights</h4><p>18 days, 63,000 GPU-hours</p></div>
</div>
<div class="observe"><i id="dr-prog" type="width:25%"></i><span class="pulse"></span></div>
<div class="element" id="dr-detail"></div>
<div class="ctrl"><button class="btn" id="dr-prev">← Prev</button><button class="btn on" id="dr-next">Subsequent →</button><button class="btn" id="dr-auto">
Auto-play</button></div>
</div>
<div class="pane" id="dr-bm">
<div class="ctrl" id="dr-bsel"></div>
<div class="bars" id="dr-bars"></div>
<div class="notice" id="dr-bnote"></div>
<div class="notice">Self-reported by RadixArk. DaisyRose scores are the upper of BF16 and FP8 runs. Competitor scores come from every developer’s reviews or cited leaderboards; some Terminal Bench 4.1 scores are 4.0 outcomes. Lacking bars weren’t reported.</div>
</div>
<div class="pane" id="dr-ef">
<div class="ctrl"><button class="btn on" data-m="mfu">MFU</button><button class="btn" data-m="tps">Throughput</button><button class="btn" data-m="gpu">GPU break up</button></div>
<div class="bars" id="dr-ebars"></div>
<div class="grid2" type="margin-top:14px">
<div class="element"><div class="massive">57%</div>much less GPU time reported for the discharge run versus the open-source baseline.</div>
<div class="element"><div class="massive">~944,000</div>sandboxes created mid-Sept to mid-Nov, peak 1,566 concurrent, 99% creation success over 9 weeks.</div>
</div>
</div>
<div class="pane" id="dr-dp">
<div class="ctrl" id="dr-psel"><button class="btn on" data-q="bf16">BF16</button><button class="btn" data-q="fp8">FP8</button><button class="btn" data-q="nvfp4">NVFP4</button><button class="btn" data-q="gguf">GGUF</button></div>
<div class="grid2" type="margin-top:12px">
<div class="element"><div class="massive" id="dr-size">~52 GB</div><span id="dr-qnote"></span></div>
<div class="element"><b>Single-GPU targets listed by RadixArk</b><div class="chips"><span class="chip">H200</span><span class="chip">B200</span><span class="chip">GB300</span><span class="chip">RTX PRO 6000</span><span class="chip">RTX 5090</span><span class="chip">DGX Spark</span></div></div>
</div>
<div class="ctrl"><button class="btn on" data-s="sg">SGLang + DSpark</button><button class="btn" data-s="vl">vLLM</button></div>
<pre id="dr-code"></pre>
<div class="notice">SGLang command abridged; see the mannequin card for the total flag set. Considering-mode sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0.</div>
</div>
<div class="ft"><span>Sources: radixark.com/weblog/daisyrose-27b · huggingface.co/RadixArk/DaisyRose-27B</span><a href="https://www.marktechpost.com/" goal="_blank" rel="noopener">© Marktechpost</a></div>
</div>
<script>
(operate(){
var R=doc.getElementById(‘mtp-drose’);
operate q(s,c)operate qa(s,c)R).querySelectorAll(s))
operate resize(){attempt{father or mother.postMessage({mtpFrame:’drose’,h:R.offsetHeight+40},’*’)}catch(e){}}
qa(‘.tab’).forEach(operate(t){t.addEventListener(‘click on’,operate(){qa(‘.tab’).forEach(operate(x){x.classList.take away(‘on’)});qa(‘.pane’).forEach(operate(x){x.classList.take away(‘on’)});t.classList.add(‘on’);q(‘#dr-‘+t.dataset.p).classList.add(‘on’);if(t.dataset.p===’bm’)drawB(cur);if(t.dataset.p===’ef’)drawE(em);setTimeout(resize,60)})});
qa(‘.tile’).forEach(operate(t){t.addEventListener(‘click on’,operate(){t.classList.toggle(‘open’);setTimeout(resize,30)})});
/* layers */
var L=q(‘#dr-layers’);for(var i=0;i<64;i++){var d=doc.createElement(‘div’);d.className=”ly “+((ipercent4===3)?’att’:’lin’);L.appendChild(d)}
operate anim(){var c=qa(‘.ly’);c.forEach(operate(x){x.classList.take away(‘lit’)});c.forEach(operate(x,i){setTimeout(operate(){x.classList.add(‘lit’)},i*28)})}
L.addEventListener(‘click on’,anim);setTimeout(anim,300);
/* coaching steps */
var S=[
‘<b>Curate tasks.</b> RadixArk and Eigent AI picked about 1,200 tasks from thousands of candidates. Each task targets a specific failure of the Qwen3.8-27B base model.’,
‘<b>Agent rollouts.</b> 128 H200s generate trajectories inside AgentENV sandboxes. Each attempt allows up to 1.5 hours, 100+ tool calls and 256k+ output tokens.’,
‘<b>Async RL update.</b> 16 H200s train with Miles in fully asynchronous mode, so rollouts and optimizer steps overlap. No SFT or distillation is used.’,
‘<b>Ship weights.</b> The run took 18 days on 144 H200s, about 63,000 GPU-hours. Result: Terminal Bench 4.1 pass@1 from 5.6 to 12.0.’];
var si=0,timer=null;operate step(n){si=(n+4)%4;qa(‘.st’,q(‘#dr-flow’)).forEach(operate(x,i){x.classList.toggle(‘on’,i<=si)});q(‘#dr-prog’).type.width=((si+1)*25)+’%’;q(‘#dr-detail’).innerHTML=S[si];resize()}
q(‘#dr-next’).onclick=operate(){step(si+1)};q(‘#dr-prev’).onclick=operate(){step(si-1)};
q(‘#dr-auto’).onclick=operate(){var b=this;if(timer){clearInterval(timer);timer=null;b.textContent=”
Auto-play”;b.classList.take away(‘on’)}else{b.textContent=”❚❚ Pause”;b.classList.add(‘on’);timer=setInterval(operate(){step(si+1)},2200)}};step(0);
/* benchmarks */
var M=[‘DaisyRose-27B’,’Qwen3.8-27B’,’Muse Glimmer-30B’,’Qwen3.6-35B-A3B’,’MiMo V2.6 Flash’,’Inkling 2 Small’,’Beam 1.0′,’Muse Spark Open’,’GPT-6 Luna’];
var B={‘Terminal Bench 4.1′:[12.0,5.6,0.5,0.0,16.1,13.0,6.4,6.6,15.2],’Terminal Bench 2.1′:[75.5,73.0,51.7,44.9,87.6,81.3,80.1,76.4,73.0],’DeepSWE v1.1′:[46.2,42.2,4.4,0.0,67.9,72.1,43.3,60.0,66.6],’SWE Bench Professional v2′:[94.7,91.3,80.4,73.3,null,97.7,null,95.0,null],’HLE (textual content solely)’:[34.3,33.9,22.0,22.2,35.1,44.1,34.7,43.9,38.5],’IFBench’:[82.4,79.5,77.0,64.4,null,null,79.6,null,null],’GPQA Diamond’:[90.7,89.2,83.5,86.0,88.7,93.0,90.6,89.0,90.5]};
var cur=”Terminal Bench 4.1″,bs=q(‘#dr-bsel’);Object.keys(B).forEach(operate(okay){var b=doc.createElement(‘button’);b.className=”btn”+(okay===cur?’ on’:”);b.textContent=okay;b.onclick=operate(){qa(‘.btn’,bs).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);cur=okay;drawB(okay)};bs.appendChild(b)});
operate drawB(okay){var v=B[k],mx=Math.max.apply(null,v.filter(operate(x){return x!==null})),h=””;M.forEach(operate(m,i){if(v[i]===null)return;h+='<div class="br’+(i===0?’ me’:i===1?’ base’:”)+’"><span class="nm">’+m+'</span><div class="bg"><div class="fl" data-w="’+(mx?v[i]/mx*100:0)+’"></div></div><span class="vl">’+v[i].toFixed(1)+'</span></div>’});q(‘#dr-bars’).innerHTML=h;setTimeout(operate(){qa(‘.fl’,q(‘#dr-bars’)).forEach(operate(f){f.type.width=f.dataset.w+’%’})},40);var d=(v[0]-v[1]).toFixed(1);q(‘#dr-bnote’).innerHTML='<b type="colour:#f5c542">DaisyRose vs base: +’+d+'</b> (‘+v[1].toFixed(1)+’ → ‘+v[0].toFixed(1)+’)’;resize()}
/* effectivity */
var E={mfu:{rows:[[‘Miles internal backend’,37,’37%’],[‘Open-source Megatron-based’,16,’16%’]],mx:37},tps:{rows:[[‘Release run (Miles)’,16.8,’16.8k’],[‘Baseline’,8.3,’8.3k’]],mx:16.8},gpu:{rows:[[‘Agent inference’,128,’128′],[‘Training’,16,’16’]],mx:128}};
var em=’mfu’;qa(‘[data-m]’).forEach(operate(b){b.onclick=operate(){qa(‘[data-m]’).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);em=b.dataset.m;drawE(em)}});
operate drawE(okay){var e=E[k],h=””;e.rows.forEach(operate(r,i){h+='<div class="br’+(i===0?’ me’:”)+’"><span class="nm">’+r[0]+'</span><div class="bg"><div class="fl" data-w="’+r[1]/e.mx*100+’"></div></div><span class="vl">’+r[2]+'</span></div>’});if(okay===’tps’)h+='<div class="notice">Tokens/s finish to finish: a 2.0x speedup.</div>’;if(okay===’gpu’)h+='<div class="notice">144 H200 GPUs whole for 18 days.</div>’;q(‘#dr-ebars’).innerHTML=h;setTimeout(operate(){qa(‘.fl’,q(‘#dr-ebars’)).forEach(operate(f){f.type.width=f.dataset.w+’%’})},40);resize()}
drawE(‘mfu’);
/* deploy */
var Qs={bf16:[‘~52 GB’,’Main repo: RadixArk/DaisyRose-27B.’],fp8:[‘~29 GB’,’RadixArk/DaisyRose-27B-FP8.’],nvfp4:[‘~23 GB’,’RadixArk/DaisyRose-27B-NVFP4. Requires a Blackwell GPU.’],gguf:[‘GGUF’,’RadixArk/DaisyRose-27B-GGUF for local runtimes. Size varies by quant.’]};
qa(‘[data-q]’).forEach(operate(b){b.onclick=operate(){qa(‘[data-q]’).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);q(‘#dr-size’).textContent=Qs[b.dataset.q][0];q(‘#dr-qnote’).textContent=Qs[b.dataset.q][1]}});q(‘#dr-qnote’).textContent=Qs.bf16[1];
var C={sg:’python3 -m sglang.launch_server n –model-path RadixArk/DaisyRose-27B n –tp-size 1 –context-length 262144 n –reasoning-parser qwen3 n –tool-call-parser qwen3_coder n –speculative-algorithm DSPARK n –speculative-draft-model-path RadixArk/Qwen3.8-27B-DSpark’,vl:’pip set up vllmnvllm serve "RadixArk/DaisyRose-27B" # OpenAI-compatible API on :8000′};
q(‘#dr-code’).textContent=C.sg;qa(‘[data-s]’).forEach(operate(b){b.onclick=operate(){qa(‘[data-s]’).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);q(‘#dr-code’).textContent=C[b.dataset.s];resize()}});
window.addEventListener(‘load’,resize);window.addEventListener(‘resize’,resize);setTimeout(resize,400);
})();
</script>
</physique></html>
“>
(operate(){var f=doc.getElementById(‘mtp-drose-frame’);window.addEventListener(‘message’,operate(e){var d=e.knowledge;if(d&&d.mtpFrame===’drose’&&d.h&&f&&e.supply===f.contentWindow){f.type.peak=d.h+’px’;}});})();
How do you deploy DaisyRose-27B?
Weights are on Hugging Face in BF16, FP8, NVFP4 and GGUF. NVFP4 requires a Blackwell GPU. RadixArk recommends SGLang with DSpark speculative decoding. That setup makes use of the Qwen3.8-27B-DSpark draft mannequin. vLLM additionally works by way of vllm serve "RadixArk/DaisyRose-27B". Recommended thinking-mode sampling is temperature 1.0, top_p 0.95 and top_k 20.
How does DaisyRose-27B examine with similar-size fashions?
| Metrics | DaisyRose-27B | Qwen3.8-27B | Muse Glimmer-30B | Qwen3.6-35B-A3B |
|---|---|---|---|---|
| Developer | RadixArk | Qwen (Alibaba) | Meta Superintelligence Lab | Qwen (Alibaba) |
| Whole params | 27.8B | 27B | ~29.6B | 35B |
| Lively params | 27.8B (dense) | 27B (dense) | ~29.6B (dense) | 3B (MoE) |
| Context | 262,144 | 262,144 (extensible to 1M) | 131,072+ | 262,144 (extensible to 1,010,000) |
| License | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Enter | Textual content, picture | Textual content, picture, video | Textual content, picture | Textual content, picture |
| {Hardware} | 1 GPU; 52 GB BF16, 23 GB NVFP4 | Not disclosed | 64 GB full precision; below 20 GB 4-bit | Not disclosed |
| Terminal Bench 2.1* | 75.5 | 73.0 | 51.7 | 44.9 |
| SWE Bench Professional v2* | 94.7 | 91.3 | 80.4 | 73.3 |
| DeepSWE v1.1* | 46.2 | 42.2 | 4.4 | 0.0 |
| HLE (textual content solely)* | 34.3 | 33.9 | 22.0 | 22.2 |
| IFBench* | 82.4 | 79.5 | 77.0 | 64.4 |
| GPQA Diamond* | 90.7 | 89.2 | 83.5 | 86.0 |
| SWE-bench Professional (official card) | Not disclosed | 61.7 | 51.2 | 49.5 |
*From RadixArk’s comparability desk, which cites every developer’s reviews or leaderboards. SWE Bench Professional v2 figures hint to arXiv 2611.16289. The final row comes from every mannequin’s personal card.
Key Takeaways
- RL solely, no SFT: 1,200 duties, 18 days, 144 H200s.
- Terminal Bench 4.1 jumps from 5.6 to 12.0.
- Matches 1 GPU: 52 GB in BF16, 23 GB in NVFP4.
- Miles hit 37% MFU, 2.0x sooner finish to finish.
- Apache 2.0, however each rating is self-reported.
Take a look at the mannequin card, the technical weblog and the Miles repo. You may examine the deployment choices and specs right here via our GPULeaderboard.com chrome extension.
Additionally, be at liberty to observe us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you possibly can be part of us on telegram as nicely.
Have to companion with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and so forth.? Join with us
The put up RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Mannequin Educated With Reinforcement Studying Solely appeared first on MarkTechPost.








