Spaces:

FINAL-Bench
/

VKIE

Running

File size: 19,810 Bytes

<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>VKIE — VIDRAFT Kernel Inference Engine</title>
<meta name="description" content="VKIE (비키) — the unified inference engine. VKAE accelerates, VKUE saves, VKIE maximizes serving. Measured, reproducible.">
<style>
  :root{
    --bg:#F5F4FA; --panel:#FFFFFF; --ink:#14121F; --muted:#6B6580; --line:#E2DEF0;
    --vkie:#7C3AED; --vkie-soft:#EDE6FB; --vkae:#E23744; --vkae-soft:#FCE7E9;
    --vkue:#1F86C9; --vkue-soft:#E3F1FB; --good:#1F8A5B;
    --mono:ui-monospace,"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{ --bg:#0E0B18; --panel:#171226; --ink:#ECE8F7; --muted:#9A93B0; --line:#2A2340;
      --vkie:#A78BFA; --vkie-soft:#221A3A; --vkae:#F87A85; --vkae-soft:#2A1518;
      --vkue:#5AB4EE; --vkue-soft:#0F2233; --good:#43D08A; }
  }
  :root[data-theme="light"]{ --bg:#F5F4FA; --panel:#FFFFFF; --ink:#14121F; --muted:#6B6580; --line:#E2DEF0;
    --vkie:#7C3AED; --vkie-soft:#EDE6FB; --vkae:#E23744; --vkae-soft:#FCE7E9; --vkue:#1F86C9; --vkue-soft:#E3F1FB; --good:#1F8A5B; }
  :root[data-theme="dark"]{ --bg:#0E0B18; --panel:#171226; --ink:#ECE8F7; --muted:#9A93B0; --line:#2A2340;
    --vkie:#A78BFA; --vkie-soft:#221A3A; --vkae:#F87A85; --vkae-soft:#2A1518; --vkue:#5AB4EE; --vkue-soft:#0F2233; --good:#43D08A; }
  *{box-sizing:border-box}
  body{margin:0;background:var(--bg);color:var(--ink);font-family:var(--sans);line-height:1.55;-webkit-font-smoothing:antialiased}
  .wrap{max-width:1080px;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(--vkie);margin:0 0 14px}
  .wordmark{font-family:var(--mono);font-weight:700;font-size:clamp(40px,9vw,74px);letter-spacing:-.01em;margin:0;line-height:.98}
  .wordmark small{color:var(--muted);font-weight:400;font-size:.26em;letter-spacing:.02em;display:block;margin-top:12px}
  .nick{color:var(--vkie)}
  .tagline{font-size:clamp(16px,2.2vw,20px);color:var(--ink);margin:18px 0 0;max-width:60ch}
  .tagline b{color:var(--vkie)}

  .axes{display:grid;grid-template-columns:repeat(auto-fit,minmax(250px,1fr));gap:14px;margin:34px 0 8px}
  .axis{background:var(--panel);border:1px solid var(--line);border-radius:16px;padding:22px 20px}
  .axis .ico{font-size:30px}
  .axis h3{margin:8px 0 2px;font-size:19px;font-family:var(--mono)}
  .axis .role{font-size:13px;color:var(--muted);margin:0 0 10px}
  .axis .big{font-family:var(--mono);font-size:clamp(20px,3vw,26px);font-weight:700;letter-spacing:-.02em}
  .axis .lab{font-size:12.5px;color:var(--muted);margin-top:4px}
  .axis.vkae{border-top:3px solid var(--vkae)} .axis.vkae h3,.axis.vkae .big{color:var(--vkae)}
  .axis.vkue{border-top:3px solid var(--vkue)} .axis.vkue h3,.axis.vkue .big{color:var(--vkue)}
  .axis.vkie{border-top:3px solid var(--vkie);background:var(--vkie-soft)} .axis.vkie h3,.axis.vkie .big{color:var(--vkie)}

  h2{font-size:13px;font-family:var(--mono);letter-spacing:.14em;text-transform:uppercase;color:var(--muted);margin:46px 0 4px;font-weight:600}
  .h2sub{font-size:14px;color:var(--muted);margin:0 0 16px}.h2sub b{color:var(--ink)}

  .ba{display:grid;grid-template-columns:1fr 1fr 1fr;gap:14px}
  @media(max-width:820px){.ba{grid-template-columns:1fr}}
  .bacard{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:18px}
  .bacard .t{font-family:var(--mono);font-weight:700;font-size:15px;margin:0 0 2px;display:flex;align-items:center;gap:8px}
  .bacard.vkae .t{color:var(--vkae)} .bacard.vkue .t{color:var(--vkue)} .bacard.vkie .t{color:var(--vkie)}
  .bacard .cond{font-size:12px;color:var(--muted);margin:0 0 14px}
  .barow{display:flex;align-items:center;justify-content:space-between;gap:10px}
  .bacol{flex:1}
  .bacol .k{font-size:11px;color:var(--muted);font-family:var(--mono);text-transform:uppercase;letter-spacing:.06em}
  .bacol .v{font-family:var(--mono);font-size:22px;font-weight:700;font-variant-numeric:tabular-nums}
  .before .v{color:var(--muted)} .after .v{color:var(--ink)}
  .arrow{font-size:20px;color:var(--muted)}
  .mult{margin-top:12px;font-family:var(--mono);font-size:13px;font-weight:700}
  .bacard.vkae .mult{color:var(--vkae)} .bacard.vkue .mult{color:var(--vkue)} .bacard.vkie .mult{color:var(--vkie)}
  .mnote{font-size:11px;color:var(--muted);margin-top:2px}

  .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:640px;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:13px 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:13px 16px;border-bottom:1px solid var(--line)}
  tbody tr:last-child td{border-bottom:none}
  .num{font-family:var(--mono);font-weight:700;font-variant-numeric:tabular-nums}
  tr.peak td{background:var(--vkie-soft)} tr.peak .num{color:var(--vkie)}
  tr.free td{background:var(--vkue-soft)} tr.free .num{color:var(--vkue)}

  .tests{display:grid;grid-template-columns:repeat(auto-fit,minmax(230px,1fr));gap:12px;margin-top:8px}
  .tcard{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:16px;text-decoration:none;color:inherit;transition:border-color .15s,transform .1s;display:block}
  .tcard:hover{border-color:var(--vkie);transform:translateY(-2px)}
  .tcard .em{font-size:20px}
  .tcard .tt{font-weight:600;font-size:14.5px;margin:6px 0 2px}
  .tcard .dd{font-size:12.5px;color:var(--muted)}
  .tcard .go{font-family:var(--mono);font-size:12px;color:var(--vkie);margin-top:8px}

  .embed{background:var(--panel);border:1px solid var(--line);border-radius:16px;padding:14px;margin-top:14px}
  .embed .eh{display:flex;align-items:center;justify-content:space-between;gap:10px;flex-wrap:wrap;margin-bottom:10px}
  .embed .eh b{font-size:14.5px}
  .loadbtn{font-family:var(--sans);font-size:13px;font-weight:700;color:#fff;background:var(--vkie);border:0;border-radius:9px;padding:9px 15px;cursor:pointer}
  .frame{width:100%;height:640px;border:1px solid var(--line);border-radius:10px;background:var(--bg)}
  .ph{height:200px;display:flex;align-items:center;justify-content:center;color:var(--muted);font-size:13px;border:1px dashed var(--line);border-radius:10px;text-align:center;padding:16px}

  .badges{display:flex;flex-wrap:wrap;gap:9px;margin-top:16px}
  .badge{font-family:var(--mono);font-size:12.5px;text-decoration:none;padding:8px 13px;border:1px solid var(--line);border-radius:999px;color:var(--muted);background:var(--panel)}
  .badge:hover{border-color:var(--vkie);color:var(--vkie)}
  .badge b{color:var(--ink)}

  .band{display:grid;grid-template-columns:repeat(auto-fit,minmax(250px,1fr));gap:12px;margin-top:8px}
  .icard{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:18px}
  .icard .ci{width:46px;height:46px;border-radius:50%;display:flex;align-items:center;justify-content:center;font-size:22px;margin-bottom:11px}
  .icard h4{margin:0 0 4px;font-size:17px;font-family:var(--mono)}
  .icard p{margin:0;font-size:13px;color:var(--muted)}

  .note{margin-top:22px;padding:15px 17px;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}
</style>
</head>
<body>
<div class="wrap">

  <p class="eyebrow">VIDRAFT · Unified Inference Engine</p>
  <h1 class="wordmark">VKIE <span class="nick">비키</span><small>VIDRAFT Kernel Inference Engine</small></h1>
  <p class="tagline"><b>VKAE</b> accelerates. <b>VKUE</b> saves. <b>VKIE</b> maximizes serving. One 34.7B model, from a datacenter GPU down to a free CPU — <b>every number measured</b>, every demo live.</p>

  <!-- PROBLEM -->
  <h2>The problem</h2>
  <p class="h2sub">Speed, cost, and scale pull against each other — win one, usually lose two.</p>
  <div class="band">
    <div class="icard"><div class="ci" style="background:var(--vkae-soft);color:var(--vkae)">⚡</div><h4>Speed</h4><p>Slow inference is unusable — but going fast normally means an expensive datacenter GPU.</p></div>
    <div class="icard"><div class="ci" style="background:var(--vkue-soft);color:var(--vkue)">💸</div><h4>Cost</h4><p>Serving a large model means a multi-GPU cluster. With GPU shortages, you often can't even buy them.</p></div>
    <div class="icard"><div class="ci" style="background:var(--vkie-soft);color:var(--vkie)">📈</div><h4>Scale</h4><p>Model and hardware get locked in — no single engine spans datacenter down to edge / on-prem.</p></div>
  </div>

  <!-- 3 AXES -->
  <div class="axes">
    <div class="axis vkae">
      <div class="ico">🏎️</div><h3>VKAE</h3><p class="role">GPU acceleration · the sports car</p>
      <div class="big">~9× faster</div><div class="lab">single-stream, 1×B200 (24 → 220 tok/s)</div>
    </div>
    <div class="axis vkue">
      <div class="ico">🚗</div><h3>VKUE</h3><p class="role">GPU savings · the compact car</p>
      <div class="big">runs on FREE CPU</div><div class="lab">34.7B at ~6-7 tok/s, zero GPU</div>
    </div>
    <div class="axis vkie">
      <div class="ico">🚄</div><h3>VKIE (비키)</h3><p class="role">accel + savings = max serving · the train</p>
      <div class="big">18,057 tok/s</div><div class="lab">1×B200 serving capacity (aggregate)</div>
    </div>
  </div>

  <!-- BEFORE / AFTER -->
  <h2>Before → After · measured</h2>
  <p class="h2sub">Each axis, on the <b>same hardware</b>, same 34.7B model — what our engine changes.</p>
  <div class="ba">
    <div class="bacard vkae">
      <div class="t">🏎️ VKAE — acceleration</div>
      <div class="cond">single-stream · 1×B200</div>
      <div class="barow">
        <div class="bacol before"><div class="k">Before</div><div class="v">24</div></div>
        <div class="arrow">→</div>
        <div class="bacol after"><div class="k">After</div><div class="v">220</div></div>
      </div>
      <div class="mult">≈ 9× faster</div>
      <div class="mnote">baseline serving → VIDRAFT optimized (tok/s)</div>
    </div>
    <div class="bacard vkue">
      <div class="t">🚗 VKUE — savings</div>
      <div class="cond">same 8 GB laptop · dense → sparse</div>
      <div class="barow">
        <div class="bacol before"><div class="k">Before</div><div class="v">5.4</div></div>
        <div class="arrow">→</div>
        <div class="bacol after"><div class="k">After</div><div class="v">20.0</div></div>
      </div>
      <div class="mult">≈ 3.7× faster</div>
      <div class="mnote">dense 32B → sparse A3B, identical hardware (tok/s)</div>
    </div>
    <div class="bacard vkie">
      <div class="t">🚄 VKIE — serving capacity</div>
      <div class="cond">1×B200 · single → optimized concurrent</div>
      <div class="barow">
        <div class="bacol before"><div class="k">Before</div><div class="v">24</div></div>
        <div class="arrow">→</div>
        <div class="bacol after"><div class="k">After</div><div class="v">18,057</div></div>
      </div>
      <div class="mult">≈ 750× serving capacity</div>
      <div class="mnote">naive single request → VKIE concurrent serving (aggregate tok/s)</div>
    </div>
  </div>

  <!-- COMPARISON -->
  <h2>Three axes at a glance</h2>
  <p class="h2sub">Same 34.7B model, same principles — three optimization targets.</p>
  <div class="tablecard"><div class="scroll"><table>
    <thead><tr><th></th><th>🏎️ VKAE</th><th>🚗 VKUE</th><th>🚄 VKIE</th></tr></thead>
    <tbody>
      <tr><td>Focus</td><td>Speed</td><td>Savings</td><td>Unified · throughput</td></tr>
      <tr><td>Best hardware</td><td>Datacenter GPU</td><td>CPU ~ small GPU</td><td>Full range</td></tr>
      <tr><td>Strength</td><td>Top single-stream speed</td><td>Lowest cost · accessibility</td><td>Max tok/s · cost-efficiency</td></tr>
      <tr><td>Measured before→after</td><td class="num">24→220 (9×)</td><td class="num">5.4→20 (3.7×)</td><td class="num">24→18,057 (750×)</td></tr>
      <tr class="peak"><td>In a phrase</td><td>fastest</td><td>cheapest</td><td>most</td></tr>
    </tbody>
  </table></div></div>

  <!-- SPECTRUM -->
  <h2>One model, the whole spectrum · measured</h2>
  <p class="h2sub"><b>Ourbox-35B-JGOS</b> — 34.7B total / ~3B active MoE. Same weights, only the hardware changes.</p>
  <div class="tablecard"><div class="scroll"><table>
    <thead><tr><th>Hardware</th><th class="r">Measured tok/s</th><th>Axis</th></tr></thead>
    <tbody>
      <tr class="peak"><td>1× B200 (datacenter)</td><td class="r num">18,057</td><td>VKIE ceiling · aggregate (256 concurrent)</td></tr>
      <tr><td>1× A10G (cloud GPU)</td><td class="r num">126</td><td>VKUE · single-stream</td></tr>
      <tr><td>8 GB gaming laptop (RTX 5060)</td><td class="r num">20.0</td><td>VKUE · 3.7× a dense 32B</td></tr>
      <tr><td>CPU-Upgrade (8 vCPU, no GPU)</td><td class="r num">~17</td><td>VKUE</td></tr>
      <tr class="free"><td>FREE CPU space (2 vCPU, no GPU)</td><td class="r num">~6-7</td><td>VKUE floor · zero cost</td></tr>
    </tbody>
  </table></div></div>
  <p class="h2sub" style="margin-top:14px">Quality holds across every tier — <b>GPQA Diamond 86.4%</b> (Ourbox-35B, maj@8) up to <b>90.9%</b> (Darwin-398B). Multimodal: Janus-Pro-1B image generation on a low-cost T4 in <b>~28 s/image</b> (fp16, 2.1× our first cut).</p>

  <!-- WHY IT MATTERS -->
  <h2>Why it matters</h2>
  <p class="h2sub">Breaking the rule that "big AI needs big money."</p>
  <div class="band">
    <div class="icard"><div class="ci" style="background:var(--vkie-soft);color:var(--vkie)">🏛️</div><h4>Sovereign AI</h4><p>Public sector, defense, healthcare, finance — data that can't touch the cloud. Frontier reasoning on an air-gapped on-prem CPU.</p></div>
    <div class="icard"><div class="ci" style="background:var(--vkue-soft);color:var(--vkue)">💰</div><h4>Cost collapse</h4><p>From a multi-hundred-thousand-dollar GPU cluster to a ~$1,600 card — or a free CPU. Entry cost drops by orders of magnitude.</p></div>
    <div class="icard"><div class="ci" style="background:var(--vkae-soft);color:var(--vkae)">🌍</div><h4>Accessibility</h4><p>Individuals, startups, SMBs, public bodies — anyone. Ready for the surge in on-device and edge demand.</p></div>
  </div>

  <!-- LIVE TESTS (reuse existing spaces) -->
  <h2>Live tests · try it yourself</h2>
  <p class="h2sub">These run on their <b>real hardware</b> (not this page). Click a card to open, or load one on screen below.</p>
  <div class="tests">
    <a class="tcard" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener"><div class="em">🟢</div><div class="tt">34.7B on one A10G</div><div class="dd">Normally an H100 job — VKUE runs it on a 24&nbsp;GB A10G, live tok/s.</div><div class="go">Open ↗</div></a>
    <a class="tcard" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-CPU" target="_blank" rel="noopener"><div class="em">🔵</div><div class="tt">CPU only · no GPU</div><div class="dd">34.7B on 8 vCPU, zero GPU.</div><div class="go">Open ↗</div></a>
    <a class="tcard" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-FreeCPU" target="_blank" rel="noopener"><div class="em">🆓</div><div class="tt">FREE CPU (~6-7 tok/s)</div><div class="dd">34.7B on HuggingFace's free CPU tier.</div><div class="go">Open ↗</div></a>
    <a class="tcard" href="https://huggingface.co/spaces/FINAL-Bench/Janus-Pro-Image-T4" target="_blank" rel="noopener"><div class="em">🖼️</div><div class="tt">Image gen (T4)</div><div class="dd">Autoregressive image on a low-cost GPU.</div><div class="go">Open ↗</div></a>
    <a class="tcard" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener"><div class="em">🏎️</div><div class="tt">VKAE — speed</div><div class="dd">Datacenter throughput leaderboard.</div><div class="go">Open ↗</div></a>
    <a class="tcard" href="https://huggingface.co/spaces/FINAL-Bench/VKUE" target="_blank" rel="noopener"><div class="em">🚗</div><div class="tt">VKUE — efficiency</div><div class="dd">Minimal-hardware leaderboard.</div><div class="go">Open ↗</div></a>
  </div>

  <div class="embed">
    <div class="eh"><b>🆓 Load VKUE on screen — 34.7B on a FREE CPU (no GPU)</b><button class="loadbtn" id="b1">▶ Load live</button></div>
    <div class="ph" id="p1">Click "Load live" to embed the free-CPU demo here (34.7B, ~6-7 tok/s, zero GPU).</div>
  </div>
  <div class="embed">
    <div class="eh"><b>🟢 Load VKUE on screen — 34.7B on one A10G GPU</b><button class="loadbtn" id="b2">▶ Load live</button></div>
    <div class="ph" id="p2">Click "Load live" to embed the A10G demo — 34.7B, which normally needs an H100 (wakes the A10G, ~1 min).</div>
  </div>

  <!-- BADGES -->
  <h2>All links</h2>
  <div class="badges">
    <a class="badge" href="https://huggingface.co/spaces/VIDraft/vkae" target="_blank" rel="noopener">🏎️ <b>VKAE</b> Speed</a>
    <a class="badge" href="https://huggingface.co/spaces/FINAL-Bench/VKUE" target="_blank" rel="noopener">🚗 <b>VKUE</b> Efficiency</a>
    <a class="badge" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-Demo" target="_blank" rel="noopener">🟢 A10G · 34.7B</a>
    <a class="badge" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-CPU" target="_blank" rel="noopener">🔵 CPU-only</a>
    <a class="badge" href="https://huggingface.co/spaces/FINAL-Bench/Ourbox-35B-VKUE-FreeCPU" target="_blank" rel="noopener">🆓 FREE CPU</a>
    <a class="badge" href="https://huggingface.co/spaces/FINAL-Bench/Janus-Pro-Image-T4" target="_blank" rel="noopener">🖼️ Image (T4)</a>
    <a class="badge" href="https://huggingface.co/FINAL-Bench/Ourbox-35B-JGOS-GGUF" target="_blank" rel="noopener">🤗 Model</a>
    <a class="badge" href="https://huggingface.co/blog/FINAL-Bench/vkue" target="_blank" rel="noopener">📝 Blog</a>
  </div>

  <div class="note">
    <b>Honest scope.</b> Every figure above is our own measurement, reproducible on the linked demos. The <i>"before"</i> baselines use standard open tooling; the <i>"after"</i> numbers come from VIDRAFT's optimized serving — the resulting speed and hardware are public, the engine internals are proprietary. VKAE/VKIE serving numbers are on a single B200; VKUE numbers are on the exact consumer/CPU hardware named. No claim that a CPU beats a GPU — VKUE's point is that the model <i>runs</i> where a GPU isn't available.
  </div>

  <footer>
    <span>FINAL-Bench · VIDRAFT</span>
    <span>VKIE (비키) · VKAE fast · VKUE anywhere · VKIE most</span>
  </footer>
</div>

<script>
  function load(btnId, phId, url){
    document.getElementById(btnId).addEventListener('click', function(){
      var ph=document.getElementById(phId);
      var f=document.createElement('iframe');
      f.className='frame'; f.src=url; f.loading='lazy';
      ph.replaceWith(f); this.disabled=true; this.textContent='Loaded';
    });
  }
  load('b1','p1','https://final-bench-ourbox-35b-vkue-freecpu.hf.space');
  load('b2','p2','https://final-bench-ourbox-35b-vkue-demo.hf.space');
</script>
</body>
</html>