Spaces:

FINAL-Bench
/

Aether-Sovereign-AI

Running on T4

File size: 11,791 Bytes

<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Aether-7B-5Attn — 100% Open-source Foundation Model: Sovereign AI</title>
<meta name="description" content="Aether-7B-5Attn is a 100% open-source sovereign foundation model: weights, training-data recipe, code, hyperparameters and full logs, all released under Apache-2.0. A 6.59B heterogeneous-attention MoE built from scratch by VIDRAFT.">
<meta name="keywords" content="open source LLM, fully open foundation model, sovereign AI, Korean LLM, open weights data code logs, Apache 2.0, heterogeneous attention, Latin square attention, mixture of experts, reproducible LLM, VIDRAFT, Aether">
<meta name="author" content="VIDRAFT">
<meta property="og:type" content="website">
<meta property="og:title" content="Aether-7B-5Attn — 100% Open-source Foundation Model: Sovereign AI">
<meta property="og:description" content="100% open-source sovereign foundation model — weights, data recipe, code, and full logs under Apache-2.0.">
<meta name="twitter:card" content="summary_large_image">
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "SoftwareApplication",
  "name": "Aether-7B-5Attn",
  "applicationCategory": "Large Language Model",
  "operatingSystem": "Any",
  "creator": {"@type": "Organization", "name": "VIDRAFT"},
  "license": "https://www.apache.org/licenses/LICENSE-2.0",
  "description": "A 100% open-source sovereign foundation model: a 6.59B heterogeneous-attention MoE (5 attention mechanisms on a 7x7 Latin square) trained from scratch on 144.2B tokens. Weights, training-data recipe, code, hyperparameters and full logs released under Apache-2.0.",
  "keywords": "open source LLM, sovereign AI, fully open foundation model, Apache-2.0, heterogeneous attention, mixture of experts"
}
</script>
<style>
  :root{--bg:#0B1020;--card:#121a30;--line:#233; --ink:#E7EEFB;--mut:#93A4C4;--acc:#5B8FF9;--good:#5AD8A6;}
  *{box-sizing:border-box}
  body{margin:0;background:radial-gradient(1200px 600px at 70% -10%,#16213f 0,#0B1020 60%);color:var(--ink);
       font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,"Noto Sans KR",sans-serif;line-height:1.6}
  .wrap{max-width:1100px;margin:0 auto;padding:0 20px}
  header{padding:56px 0 28px}
  .badge{display:inline-block;font-size:12px;letter-spacing:.14em;color:var(--good);border:1px solid #2b6b52;
         border-radius:999px;padding:5px 12px;text-transform:uppercase}
  h1{font-size:clamp(28px,5vw,46px);line-height:1.15;margin:16px 0 10px;font-weight:800}
  h1 .g{background:linear-gradient(90deg,#7FB2FF,#5AD8A6);-webkit-background-clip:text;background-clip:text;color:transparent}
  .sub{color:var(--mut);font-size:18px;max-width:760px}
  .cta{margin:22px 0 4px;display:flex;gap:12px;flex-wrap:wrap}
  .btn{display:inline-block;padding:11px 18px;border-radius:10px;text-decoration:none;font-weight:600}
  .btn.p{background:var(--acc);color:#06122b}
  .btn.o{border:1px solid #2c3c5e;color:var(--ink)}
  .grid{display:grid;grid-template-columns:1.15fr .85fr;gap:22px;margin:26px 0}
  @media(max-width:860px){.grid{grid-template-columns:1fr}}
  .card{background:var(--card);border:1px solid #1f2b47;border-radius:16px;padding:20px}
  .card h2{margin:2px 0 10px;font-size:19px}
  .demo textarea{width:100%;min-height:110px;background:#0c142a;color:var(--ink);border:1px solid #26324a;
                 border-radius:10px;padding:12px;font-size:15px;resize:vertical;font-family:inherit}
  .demo .row{display:flex;gap:10px;align-items:center;margin-top:10px;flex-wrap:wrap}
  .demo button{background:var(--good);color:#06251a;border:0;border-radius:10px;padding:11px 18px;font-weight:700;cursor:pointer}
  .demo button:disabled{opacity:.5;cursor:not-allowed}
  .out{margin-top:12px;background:#0a1226;border:1px solid #23304d;border-radius:10px;padding:14px;white-space:pre-wrap;
       min-height:70px;color:#D8E4FA}
  .muted{color:var(--mut);font-size:13px}
  ul.feat{list-style:none;padding:0;margin:6px 0}
  ul.feat li{padding:7px 0;border-bottom:1px solid #18223c}
  ul.feat li b{color:var(--good)}
  table{width:100%;border-collapse:collapse;font-size:14px;margin-top:6px}
  th,td{text-align:left;padding:8px 10px;border-bottom:1px solid #1a2540}
  th{color:var(--mut);font-weight:600}
  .qa h3{margin:16px 0 4px;font-size:16px;color:#BFE3FF}
  .qa p{margin:0 0 10px;color:var(--mut)}
  footer{color:#6B7C99;font-size:13px;padding:30px 0 50px;border-top:1px solid #17203a;margin-top:20px}
  a{color:#7FB2FF}
</style>
</head>
<body>
<div class="wrap">
  <header>
    <span class="badge">Apache-2.0 · Fully Open · Sovereign AI</span>
    <h1>Aether-7B-5Attn — <span class="g">100% Open-source</span> Foundation Model</h1>
    <p class="sub">Not "open weights." Genuinely open — weights, training-data recipe, code, every hyperparameter, and the complete logs, all released under Apache-2.0. A 6.59B heterogeneous-attention MoE, built from scratch by <b>VIDRAFT</b>.</p>
    <div class="cta">
      <a class="btn p" href="#try">Try the model</a>
      <a class="btn o" href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn" target="_blank" rel="noopener">Base model ↗</a>
      <a class="btn o" href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn-it" target="_blank" rel="noopener">Instruct model ↗</a>
    </div>
  </header>

  <div class="grid">
    <div class="card demo" id="try">
      <h2>🌏 Try Aether (live)</h2>
      <p class="muted">Answers here come from <strong><a href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn-it">Aether-7B-5Attn-it</a></strong>, the instruction-tuned model — <em>not</em> the base. The base (<a href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn">Aether-7B-5Attn</a>) is the reproducible artifact this release is about; it is not instruction-tuned and would not answer questions like this.</p>
      <textarea id="prompt" placeholder="Ask something, or start a sentence for the model to continue...">In one sentence, what does sovereign open-source AI mean?</textarea>
      <div class="row">
        <button id="go">Generate</button>
        <span class="muted" id="stat">Model loads on first request.</span>
      </div>
      <div class="out" id="out">—</div>
      <p class="muted">Serving <code>Aether-7B-5Attn-it</code> (6.59B MoE) on a single T4, at batch_size=1. This build ships no KV cache, so generation is slow and capped — tokens stream in as they are produced.</p>
    </div>

    <div class="card">
      <h2>Why it matters</h2>
      <ul class="feat">
        <li><b>100% open</b> — weights · data recipe · code · logs</li>
        <li><b>Reproducible</b> — data rebuilds byte-for-byte from public sources</li>
        <li><b>Self-designed</b> — 5 attention mechanisms on a 7×7 Latin square</li>
        <li><b>Sovereign</b> — rebuild your own foundation model from this repo alone</li>
        <li><b>Apache-2.0</b> — commercial use, modification, redistribution</li>
      </ul>
    </div>
  </div>

  <div class="card">
    <h2>What is Aether-7B-5Attn?</h2>
    <p class="muted">A 6.59-billion-parameter Mixture-of-Experts language model (~2.98B active per token) whose 49 layers place <b>five distinct attention mechanisms</b> on a <b>7×7 Latin square</b>, so each mechanism appears exactly once at every depth. Trained from scratch on <b>144.2B tokens</b> (≈37.8% math, 21.6% Korean, 21.6% English, 13.5% code) and released fully open under Apache-2.0.</p>
    <table>
      <tr><th>Total / active params</th><td>6.59B / ~2.98B</td><th>Layers</th><td>49 (7×7 Latin square)</td></tr>
      <tr><th>Experts</th><td>25, top-7 + 1 shared</td><th>Context</th><td>4096</td></tr>
      <tr><th>Tokens</th><td>144.2B</td><th>License</th><td>Apache-2.0</td></tr>
    </table>
  </div>

  <div class="card">
    <h2>The Aether family</h2>
    <p class="muted">One architecture line, released at different levels of openness.</p>
    <ul>
      <li><b><a href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn">Aether-7B-5Attn</a></b> — base, <b>fully open</b>: weights + data recipe + training code + complete logs + intermediate checkpoints.</li>
      <li><b><a href="https://huggingface.co/FINAL-Bench/Aether-7B-5Attn-it">Aether-7B-5Attn-it</a></b> — instruction-tuned. This is the model answering in the demo above.</li>
      <li><b><a href="https://huggingface.co/FINAL-Bench/Aether-7B-7Attn-base">Aether-7B-7Attn-base</a></b> — open weights.</li>
      <li><b><a href="https://huggingface.co/FINAL-Bench/Aether-6B-11Attn">Aether-6B-11Attn</a></b> — open weights. 121 layers place <b>11 heterogeneous sequence-mixing mechanisms</b> — attention, state-space (Mamba2), convolutional (Hyena) and linear families — on an <b>11×11 Latin square</b>.</li>
      <li><a href="https://huggingface.co/datasets/FINAL-Bench/Aether-7B-5Attn-checkpoints">Intermediate checkpoints</a> (110k · 115k · 162k) · <a href="https://huggingface.co/blog/FINAL-Bench/opensource-llm">Blog write-up</a> · <a href="https://huggingface.co/collections/FINAL-Bench/aether-foundation-model">Collection</a></li>
    </ul>
  </div>

  <div class="card qa">
    <h2>Answers</h2>
    <h3>Is Aether open source?</h3>
    <p>Yes — 100% open. Weights + training-data recipe + code + all hyperparameters + complete logs + full architecture source, under Apache-2.0. This is the "fully open" standard, not "open weights only."</p>
    <h3>Who built it?</h3>
    <p>VIDRAFT (주식회사 비드래프트), a Korean AI startup — the only single startup among the world's sovereign fully-open foundation models (alongside OLMo, Soofi, Apertus, LLM-jp, Viking).</p>
    <h3>What is unique about the architecture?</h3>
    <p>It is the only sovereign fully-open model with a self-designed attention layout: five mechanisms (full, differential, sliding, NSA, hybrid) on a 7×7 Latin square — a control that removes depth bias.</p>
    <h3>Can I reproduce the training data?</h3>
    <p>Yes. Every source is a public repository; the released recipe reproduces the training binaries byte-for-byte.</p>
  </div>

  <footer>
    Built by VIDRAFT (주식회사 비드래프트) · Contact: arxivgpt@gmail.com · License: Apache-2.0<br>
    Keywords: open source LLM · fully open foundation model · sovereign AI · Korean LLM · Apache-2.0 · heterogeneous attention · mixture of experts · reproducible LLM
  </footer>
</div>

<script>
const out=document.getElementById('out'), btn=document.getElementById('go'), stat=document.getElementById('stat');
async function gen(){
  const prompt=document.getElementById('prompt').value.trim();
  if(!prompt){return}
  btn.disabled=true; out.textContent=''; stat.textContent='generating (model may be loading on first call)…';
  const t0=performance.now();
  try{
    const r=await fetch('/api/stream',{method:'POST',headers:{'Content-Type':'application/json'},
      body:JSON.stringify({prompt:prompt,max_new_tokens:200,temperature:0.7})});
    if(!r.body){ out.textContent=await r.text(); stat.textContent=''; btn.disabled=false; return; }
    const reader=r.body.getReader(), dec=new TextDecoder();
    let text='';
    while(true){
      const {done,value}=await reader.read();
      if(done) break;
      text+=dec.decode(value,{stream:true});
      out.textContent=text;
      stat.textContent=`streaming… ${((performance.now()-t0)/1000).toFixed(1)}s`;
    }
    const secs=((performance.now()-t0)/1000).toFixed(1);
    out.textContent=text||'(empty)';
    stat.textContent = r.status>=400 ? '' : `done · ${secs}s · live token streaming`;
  }catch(e){ out.textContent='Request failed: '+e; stat.textContent=''; }
  btn.disabled=false;
}
btn.addEventListener('click',gen);
</script>
</body>
</html>