Hugging Face's logo

Spaces:
FINAL-Bench
/
VKUE
Running

App Files Community
VKUE / index.html
SeaWolf-AI's picture
SeaWolf-AI
Add live A10G-vs-CPU demo links + measured 126/17 tok/s
699674b verified
Raw
History Blame Contribute Delete
17.4 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>VKUE — VIDRAFT Kernel Ubiquitous Engine</title>
<meta name="description" content="VKUE Efficiency Leaderboard — serve frontier-class models on minimal hardware.">
</head>
<body>
<style>
:root{
--ground:#F4F7F7; --panel:#FFFFFF; --ink:#0B1418; --muted:#5B6B70;
--line:#D7E0E1; --accent:#0E8C9E; --accent-soft:#E2F4F6; --hero:#B0791A;
--hero-soft:#FBF1DE; --good:#1F8A5B; --good-soft:#E4F3EC; --star:#0E8C9E;
--mono:ui-monospace,"SFMono-Regular","Cascadia Code","Consolas",monospace;
--sans:-apple-system,BlinkMacSystemFont,"Segoe UI","Apple SD Gothic Neo","Malgun Gothic",Roboto,sans-serif;
}
@media (prefers-color-scheme:dark){
:root{
--ground:#0B1418; --panel:#111E23; --ink:#E6EFF0; --muted:#8CA1A6;
--line:#22343A; --accent:#22B8CF; --accent-soft:#0F2A31; --hero:#E0A83E;
--hero-soft:#2A2113; --good:#43D08A; --good-soft:#10281D; --star:#22B8CF;
}
}
:root[data-theme="light"]{
--ground:#F4F7F7; --panel:#FFFFFF; --ink:#0B1418; --muted:#5B6B70;
--line:#D7E0E1; --accent:#0E8C9E; --accent-soft:#E2F4F6; --hero:#B0791A;
--hero-soft:#FBF1DE; --good:#1F8A5B; --good-soft:#E4F3EC; --star:#0E8C9E;
}
:root[data-theme="dark"]{
--ground:#0B1418; --panel:#111E23; --ink:#E6EFF0; --muted:#8CA1A6;
--line:#22343A; --accent:#22B8CF; --accent-soft:#0F2A31; --hero:#E0A83E;
--hero-soft:#2A2113; --good:#43D08A; --good-soft:#10281D; --star:#22B8CF;
}
*{box-sizing:border-box}
body{margin:0;background:var(--ground);color:var(--ink);font-family:var(--sans);
line-height:1.5;-webkit-font-smoothing:antialiased}
.wrap{max-width:1060px;margin:0 auto;padding:clamp(20px,4vw,52px) clamp(16px,4vw,40px)}
.eyebrow{font-family:var(--mono);font-size:12px;letter-spacing:.22em;text-transform:uppercase;
color:var(--accent);margin:0 0 14px}
.wordmark{font-family:var(--mono);font-weight:700;font-size:clamp(38px,8vw,68px);
letter-spacing:-.01em;margin:0;line-height:.98;text-wrap:balance}
.wordmark small{color:var(--muted);font-weight:400;font-size:.30em;letter-spacing:.02em;display:block;margin-top:12px}
.tagline{font-size:clamp(17px,2.4vw,22px);color:var(--ink);margin:18px 0 0;max-width:52ch;text-wrap:balance}
.tagline b{color:var(--accent)}
.sibs{display:flex;flex-wrap:wrap;gap:10px;margin-top:22px}
.chip{font-family:var(--mono);font-size:12.5px;padding:7px 13px;border:1px solid var(--line);
border-radius:999px;color:var(--muted);background:var(--panel)}
.chip b{color:var(--ink)}
.chip.now{border-color:var(--accent);color:var(--accent)}
.chip.link{text-decoration:none;cursor:pointer;transition:border-color .15s,color .15s}
.chip.link:hover{border-color:var(--accent);color:var(--accent)}
.card a.tlink{text-decoration:none}
.card a.tlink:hover{text-decoration:underline}
.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(198px,1fr));gap:14px;margin:36px 0 8px}
.stat{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:20px 18px}
.stat .n{font-family:var(--mono);font-size:clamp(23px,3.8vw,32px);font-weight:700;letter-spacing:-.02em;
color:var(--accent);font-variant-numeric:tabular-nums}
.stat.hero .n{color:var(--hero)}
.stat .l{font-size:13.5px;color:var(--muted);margin-top:7px}
.stat .l b{color:var(--ink);font-weight:600}
h2{font-size:13px;font-family:var(--mono);letter-spacing:.14em;text-transform:uppercase;
color:var(--muted);margin:44px 0 4px;font-weight:600}
.h2sub{font-size:14px;color:var(--muted);margin:0 0 16px}
.h2sub b{color:var(--ink)}
.tablecard{background:var(--panel);border:1px solid var(--line);border-radius:16px;overflow:hidden}
.scroll{overflow-x:auto}
table{border-collapse:collapse;width:100%;min-width:800px;font-size:14px}
thead th{font-family:var(--mono);font-size:11px;letter-spacing:.08em;text-transform:uppercase;
color:var(--muted);text-align:left;padding:14px 16px;border-bottom:1px solid var(--line);font-weight:600;white-space:nowrap}
thead th.r,tbody td.r{text-align:right}
tbody td{padding:14px 16px;border-bottom:1px solid var(--line);vertical-align:middle}
tbody tr:last-child td{border-bottom:none}
.mname{font-weight:600;font-size:14.5px;display:flex;align-items:center;gap:8px;flex-wrap:wrap}
.mname .star{color:var(--star)}
.lineage{font-size:11.5px;color:var(--muted);margin-top:3px;font-family:var(--mono)}
.gpqa{font-family:var(--mono);font-weight:700;font-size:16px;font-variant-numeric:tabular-nums}
.method{font-size:10.5px;color:var(--muted);margin-top:2px;letter-spacing:.01em}
.size{font-family:var(--mono);font-size:12.5px;color:var(--muted);font-variant-numeric:tabular-nums;white-space:nowrap}
.hw{font-size:12.5px;color:var(--muted);white-space:nowrap}
.vkue{font-family:var(--mono);font-weight:600;font-size:13px;color:var(--ink);white-space:nowrap}
.pill{display:inline-block;font-family:var(--mono);font-size:11px;padding:3px 9px;border-radius:999px;
background:var(--accent-soft);color:var(--accent);border:1px solid transparent;white-space:nowrap}
.pill.cpu{background:var(--good-soft);color:var(--good)}
tr.hero td{background:var(--hero-soft)}
tr.hero .gpqa,tr.hero .vkue{color:var(--hero)}
tr.hero .pill{background:transparent;border-color:var(--hero);color:var(--hero)}
tr.base td{opacity:.72}
.tag-hero{font-family:var(--mono);font-size:10px;padding:2px 7px;border-radius:5px;background:var(--hero);
color:#0B1418;font-weight:700;letter-spacing:.04em}
.legend{display:grid;grid-template-columns:repeat(auto-fit,minmax(240px,1fr));gap:10px 24px;
margin-top:20px;font-size:12.5px;color:var(--muted)}
.legend div b{color:var(--ink);font-weight:600}
.legend .k{font-family:var(--mono);color:var(--accent)}
.cols{display:grid;grid-template-columns:1fr 1fr;gap:14px;margin-top:16px}
.card{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:22px}
.card .t{font-family:var(--mono);font-weight:700;font-size:15px;margin:0 0 4px}
.card.a .t{color:var(--hero)} .card.b .t{color:var(--accent)}
.card p{margin:8px 0 0;font-size:13.5px;color:var(--muted)}
.card p b{color:var(--ink)}
.note{margin-top:16px;padding:14px 16px;border:1px dashed var(--line);border-radius:12px;
font-size:12.5px;color:var(--muted);background:var(--panel)}
.note b{color:var(--ink)}
footer{margin-top:40px;padding-top:20px;border-top:1px solid var(--line);
font-size:12px;color:var(--muted);font-family:var(--mono);display:flex;justify-content:space-between;flex-wrap:wrap;gap:8px}
@media (max-width:680px){.stats,.cols{grid-template-columns:1fr}}
</style>
<div class="wrap">
<p class="eyebrow">VIDRAFT · Inference Engine</p>
<h1 class="wordmark">VKUE<small>VIDRAFT Kernel Ubiquitous Engine</small></h1>
<p class="tagline">No GPU? Runs anyway. <b>Everywhere.</b> The ubiquity engine that serves frontier-class models on minimal hardware.</p>
<div class="sibs">
<a class="chip link" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener"><b>VKAE</b> — Speed &#8599;</a>
<span class="chip now"><b>VKUE</b> — Ubiquity · Anywhere</span>
<a class="chip link" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener"><b>&#9654; Live: GPU vs CPU</b> &#8599;</a>
<a class="chip link" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-CPU" target="_blank" rel="noopener"><b>&#9654; Live: CPU-only</b> &#8599;</a>
</div>
<div class="stats">
<div class="stat hero">
<div class="n">18,057 tok/s</div>
<div class="l">Ourbox-35B on a <b>single B200</b> &mdash; VKAE speed ceiling. <b>Measured</b>, aggregate.</div>
</div>
<div class="stat">
<div class="n">20 tok/s</div>
<div class="l"><b>Same model</b> on an <b>8&nbsp;GB laptop</b> &mdash; VKUE ubiquity floor. <b>Measured</b>, 3.7&times; a dense 32B.</div>
</div>
<div class="stat">
<div class="n">744B &rarr; 1 GPU</div>
<div class="l">Even <b>GLM-5.2</b>-class giants on a single <b>H100</b> (normally 6&ndash;8).</div>
</div>
<div class="stat">
<div class="n">90.9%</div>
<div class="l">Darwin-398B-JGOS top <b>GPQA Diamond</b> (greedy, no test-time engine).</div>
</div>
</div>
<h2>VKUE Efficiency Leaderboard</h2>
<p class="h2sub">How little hardware a given quality needs &mdash; <b>27B and up</b>. GPQA values are each model card&rsquo;s posted figure (method noted).</p>
<div class="tablecard"><div class="scroll">
<table>
<thead>
<tr>
<th>Model</th>
<th class="r">GPQA Diamond</th>
<th class="r">Size (total / active)</th>
<th>Standard HW</th>
<th>VKUE Min HW</th>
</tr>
</thead>
<tbody>
<tr class="hero">
<td><div class="mname">GLM-5.2 <span class="tag-hero">SYMBOLIC</span></div><div class="lineage">Z.ai · 744B MoE · frontier</div></td>
<td class="r"><div class="gpqa">frontier</div><div class="method">official figure &mdash; citation pending</div></td>
<td class="r"><span class="size">744B / 40B</span></td>
<td><span class="hw">H100 &times;6&ndash;8 (INT4)</span></td>
<td><span class="vkue">H100 &times;1</span> <span class="pill">&asymp; 1/6&ndash;1/8</span><div class="method">capacity verified · speed pending</div></td>
</tr>
<tr>
<td><div class="mname">Darwin-398B-JGOS</div><div class="lineage">Darwin V9 · 397B MoE</div></td>
<td class="r"><div class="gpqa">90.9</div><div class="method">greedy · no engine</div></td>
<td class="r"><span class="size">397B / 17B</span></td>
<td><span class="hw">pure GPU &times;3&ndash;4 (INT4)</span></td>
<td><span class="vkue">1 GPU + RAM</span> <span class="pill">available now</span></td>
</tr>
<tr>
<td><div class="mname">Darwin-28B-REASON</div><div class="lineage">Darwin · DELPHI engine</div></td>
<td class="r"><div class="gpqa">89.39</div><div class="method">maj · DELPHI engine</div></td>
<td class="r"><span class="size">26.9B</span></td>
<td><span class="hw">1 GPU</span></td>
<td><span class="vkue">24GB &times;1</span> <span class="pill cpu">CPU-capable</span></td>
</tr>
<tr>
<td><div class="mname">Darwin-28B-Opus</div><div class="lineage">Darwin · 28B</div></td>
<td class="r"><div class="gpqa">88.89</div><div class="method">posted figure</div></td>
<td class="r"><span class="size">~28B</span></td>
<td><span class="hw">1 GPU</span></td>
<td><span class="vkue">24GB &times;1</span> <span class="pill cpu">CPU-capable</span></td>
</tr>
<tr>
<td><div class="mname"><span class="star">&#9733;</span> Darwin-36B-Opus</div><div class="lineage">base &larr; Qwen3.6-35B-A3B</div></td>
<td class="r"><div class="gpqa">88.4</div><div class="method">GPQA Diamond</div></td>
<td class="r"><span class="size">36B / 3B</span></td>
<td><span class="hw">1 GPU (FP8)</span></td>
<td><span class="vkue">consumer 24GB &times;1</span> <span class="pill cpu">CPU-capable</span></td>
</tr>
<tr>
<td><div class="mname">Darwin-27B-Opus</div><div class="lineage">Qwen3.5 · 27.8B</div></td>
<td class="r"><div class="gpqa">86.9</div><div class="method">posted figure</div></td>
<td class="r"><span class="size">27.8B</span></td>
<td><span class="hw">1 GPU</span></td>
<td><span class="vkue">24GB &times;1</span> <span class="pill cpu">CPU-capable</span></td>
</tr>
<tr class="hero">
<td><div class="mname"><span class="star">&#9733;</span> Ourbox-35B-JGOS <span class="tag-hero">MEASURED</span></div><div class="lineage">Darwin-36B derivative · <a href="https://huggingface.co/FINAL-Bench/Ourbox-35B-JGOS-GGUF" style="color:inherit">GGUF ↗</a></div></td>
<td class="r"><div class="gpqa">86.4 / 70.7</div><div class="method">maj@8 / greedy (measured)</div></td>
<td class="r"><span class="size">35B / 3B</span></td>
<td><span class="hw">1 GPU (FP8)</span></td>
<td><span class="vkue">A10G 126 &middot; CPU 17 tok/s</span> <a class="pill cpu" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener" style="text-decoration:none">live demo &#8599;</a><div class="method">live measured &middot; also 8GB laptop 20 tok/s (3.7&times; dense 32B)</div></td>
</tr>
<tr>
<td><div class="mname">Darwin-31B-Opus</div><div class="lineage">base &larr; Gemma4-31B · Korean #1</div></td>
<td class="r"><div class="gpqa">85.9</div><div class="method">posted figure</div></td>
<td class="r"><span class="size">32.7B</span></td>
<td><span class="hw">1 GPU</span></td>
<td><span class="vkue">consumer 24GB &times;1</span> <span class="pill cpu">CPU-capable</span></td>
</tr>
<tr class="base">
<td><div class="mname">Qwen3.6-35B-A3B</div><div class="lineage">base model · Darwin-36B parent</div></td>
<td class="r"><div class="gpqa">&mdash;</div><div class="method">official figure &mdash; pending</div></td>
<td class="r"><span class="size">35B / 3B</span></td>
<td><span class="hw">1 GPU (FP8)</span></td>
<td><span class="vkue">consumer 24GB &times;1</span></td>
</tr>
<tr class="base">
<td><div class="mname">Gemma4-31B</div><div class="lineage">base model · Darwin-31B parent</div></td>
<td class="r"><div class="gpqa">&mdash;</div><div class="method">official figure &mdash; pending</div></td>
<td class="r"><span class="size">31B</span></td>
<td><span class="hw">1 GPU</span></td>
<td><span class="vkue">consumer 24GB &times;1</span></td>
</tr>
</tbody>
</table>
</div></div>
<div class="legend">
<div><span class="k">Method labels</span> &middot; <b>greedy</b> = single pass, <b>maj@8</b> = majority of 8 samples, <b>DELPHI</b> = our test-time engine. Different conditions &mdash; not directly comparable.</div>
<div><span class="k">VKUE Min HW</span> &middot; INT4, held fully in VRAM or served with little / no GPU. <b>CPU-capable</b> = runs without a GPU (slower) thanks to ~3B active params.</div>
<div><span class="k">Throughput (tok/s)</span> &middot; Ourbox-35B measured: <b>18,057</b> aggregate on B200 (<a class="tlink" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener" style="color:var(--accent)">VKAE &#8599;</a>) / <b>126</b> on one A10G &amp; <b>17</b> on a GPU-less CPU box (<a class="tlink" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener" style="color:var(--accent)">live demo &#8599;</a>) / <b>20</b> on an 8GB laptop. All single-stream, generate-format&ndash;gated.</div>
<div><span class="k">Base models</span> &middot; Qwen3.6-35B / Gemma4-31B are our bases; their GPQA figures will be cited from official sources.</div>
</div>
<h2>Engine Siblings</h2>
<div class="cols">
<div class="card a">
<p class="t"><a class="tlink" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener" style="color:inherit">VKAE &mdash; Speed &#8599;</a></p>
<p><b>How fast.</b> VIDRAFT-optimized serving pushes throughput to the ceiling &mdash; Ourbox-35B <b>measured at 18,057 tok/s</b> aggregate on a single <b>B200</b>. <a class="tlink" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener" style="color:var(--accent)">Open the Speed leaderboard &#8599;</a></p>
</div>
<div class="card b">
<p class="t"><a class="tlink" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener" style="color:inherit">VKUE &mdash; Ubiquity &#8599;</a></p>
<p><b>How low the bar.</b> Same 34.7B model, minimal hardware &mdash; <b>live measured 126 tok/s on one A10G</b> vs <b>17 tok/s on a GPU-less CPU box</b> (identical weights, only the hardware differs), and <b>20 tok/s</b> on an 8&nbsp;GB laptop. If you have a card it&rsquo;s fast; if you don&rsquo;t, it still runs. <a class="tlink" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener" style="color:var(--accent)">Try the live GPU&nbsp;vs&nbsp;CPU demo &#8599;</a></p>
</div>
</div>
<div class="note">
<b>What&rsquo;s open.</b> Quantized models, benchmark figures, and open-tool run instructions are public &mdash; anyone can reproduce the baseline.
VKUE&rsquo;s engine internals are proprietary; this leaderboard reports only the resulting hardware, never the how.
</div>
<footer>
<span>FINAL-Bench &middot; VIDRAFT</span>
<span>2026 &middot; VKUE v1 preview</span>
</footer>
</div>
</body>
</html>