Model card
From the Hugging Face model README
<style>
.co-wrap {
font-family: "Iowan Old Style", "Palatino Linotype", Palatino, Georgia, serif;
max-width: 860px;
margin: 0 auto;
color: #E8E4DA;
background: #14161C;
padding: 40px 44px 48px;
border-radius: 4px;
line-height: 1.65;
}
.co-wrap h1, .co-wrap h2 { font-weight: 600; letter-spacing: -0.01em; }
.co-name {
font-size: 2.3em;
margin: 0 0 2px;
color: #F5F2EA;
line-height: 1.1;
}
.co-sub { color: #8E9AA8; font-size: 1.02em; margin: 0 0 30px; font-style: italic; }
.co-wrap h2 {
font-size: 1.22em;
color: #F5F2EA;
margin: 38px 0 12px;
padding-bottom: 7px;
border-bottom: 1px solid #2A2F39;
}
.co-wrap p, .co-wrap li { color: #C9C4B8; }
.co-h3 { font-size: 1.02em; color: #E8E4DA; font-weight: 600; margin: 24px 0 8px; }
.co-fig { margin: 4px 0 6px; }
.co-cap { color: #7E8794; font-size: 0.85em; margin: 0 0 8px; }
.co-key { display: flex; gap: 20px; flex-wrap: wrap; font-size: 0.88em; margin-bottom: 4px; }
.co-key span { display: flex; align-items: center; gap: 7px; color: #A9B2BD; }
.co-sw { width: 11px; height: 11px; border-radius: 2px; display: inline-block; }
.co-wrap table { border-collapse: collapse; width: 100%; font-size: 0.92em; margin: 6px 0 4px; }
.co-wrap th {
text-align: left; color: #8E9AA8; font-weight: 500;
border-bottom: 1px solid #2A2F39; padding: 7px 10px 7px 0;
}
.co-wrap td { padding: 7px 10px 7px 0; border-bottom: 1px solid #1F242C; color: #C9C4B8; }
.co-wrap td:first-child, .co-wrap th:first-child { color: #E8E4DA; }
.co-num { font-variant-numeric: tabular-nums; }
.co-wrap pre {
background: #0E1015; border: 1px solid #262B34; border-radius: 3px;
padding: 15px 17px; overflow-x: auto; font-size: 0.83em; line-height: 1.5;
}
.co-wrap code { font-family: "SF Mono", Menlo, Consolas, monospace; color: #9FBFB0; }
.co-wrap a { color: #8FA8DB; text-decoration: none; border-bottom: 1px solid #33415C; }
.co-wrap a:hover { border-bottom-color: #8FA8DB; }
.co-flag {
border-left: 2px solid #B4763A; padding: 2px 0 2px 16px; margin: 18px 0;
}
.co-flag p { color: #D6C4AC; margin: 0 0 8px; }
.co-foot { color: #6E7784; font-size: 0.85em; margin-top: 34px; }
/* cover art: shown whole at 2:3, bled into the page rather than framed */
.co-hero { position: relative; margin: 0 0 22px; }
.co-cover {
width: 380px; max-width: 86%; margin: 0 auto; position: relative;
aspect-ratio: 2 / 3;
}
.co-cover::before {
content: ""; position: absolute; inset: -12% -16% -9%;
background:
radial-gradient(ellipse 46% 40% at 28% 26%, rgba(214,150,68,0.17), transparent 68%),
radial-gradient(ellipse 58% 48% at 60% 62%, rgba(127,119,221,0.13), transparent 72%);
filter: blur(22px); z-index: 0;
}
.co-cover img {
position: relative; z-index: 1;
width: 100%; height: 100%; object-fit: contain; display: block;
border-radius: 47% 53% 44% 56% / 7% 7% 5% 5%;
-webkit-mask-image: radial-gradient(ellipse 84% 86% at 50% 44%, #000 62%, rgba(0,0,0,0.72) 84%, transparent 99%);
mask-image: radial-gradient(ellipse 84% 86% at 50% 44%, #000 62%, rgba(0,0,0,0.72) 84%, transparent 99%);
}
.co-cover::after {
content: ""; position: absolute; inset: 0; z-index: 2; pointer-events: none;
border-radius: 47% 53% 44% 56% / 7% 7% 5% 5%;
background:
linear-gradient(106deg, transparent 40%, rgba(226,232,245,0.05) 47%, transparent 53%),
radial-gradient(ellipse 80% 84% at 50% 44%, transparent 54%, rgba(20,22,28,0.62) 88%, rgba(20,22,28,0.98) 100%),
linear-gradient(to bottom, rgba(20,22,28,0.80) 0%, transparent 9%, transparent 74%, rgba(20,22,28,0.92) 98%);
}
/* pointer to the successor quant */
.co-next {
border: 1px solid #33415C; border-left: 3px solid #8FA8DB;
background: rgba(143,168,219,0.07);
border-radius: 3px; padding: 13px 18px; margin: 0 0 24px;
font-size: 0.94em;
}
.co-next p { margin: 0; }
.co-next .co-next-label {
display: block; font-family: monospace; font-size: 0.74em;
letter-spacing: 0.15em; text-transform: uppercase;
color: #8FA8DB; margin-bottom: 5px;
}
.co-next a { font-weight: 600; }
@media (prefers-color-scheme: light) {
.co-wrap { background: #FBFAF7; color: #23282F; }
.co-name, .co-wrap h2, .co-h3 { color: #14171C; }
.co-wrap p, .co-wrap li, .co-wrap td { color: #3C434C; }
.co-wrap th, .co-sub, .co-cap { color: #6B7480; }
.co-wrap h2 { border-bottom-color: #DDD9D0; }
.co-wrap td { border-bottom-color: #EAE7E0; }
.co-wrap pre { background: #F2F0EA; border-color: #DDD9D0; }
.co-wrap code { color: #2C5F49; }
.co-wrap a { color: #3A5A9B; border-bottom-color: #B8C4DC; }
.co-cover::after {
background:
linear-gradient(106deg, transparent 40%, rgba(255,255,255,0.10) 47%, transparent 53%),
radial-gradient(ellipse 80% 84% at 50% 44%, transparent 54%, rgba(251,250,247,0.62) 88%, rgba(251,250,247,0.98) 100%),
linear-gradient(to bottom, rgba(251,250,247,0.80) 0%, transparent 9%, transparent 74%, rgba(251,250,247,0.92) 98%);
}
.co-next { border-color: #B8C4DC; border-left-color: #3A5A9B; background: rgba(58,90,155,0.06); }
.co-next .co-next-label { color: #3A5A9B; }
}
</style>
<div class="co-wrap">
<div class="co-hero">
<div class="co-cover">
<img src="cover.jpg" alt="">
</div>
</div>
<h1 class="co-name">Omega-Sapphira-Joyous-L3.3-70B-v1.0</h1>
<p class="co-sub">EXL3 · 4.25 bits per weight · 8-bit head</p>
<div class="co-next">
<span class="co-next-label">Newer cut available</span>
<p><a href="https://huggingface.co/cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.1_EXL3_4.25bpw_H8">v1.1 at the same bitrate</a> — same graft with the Joyous share cut back sharply, after the amounts used here proved too heavy. Same size, same VRAM fit.</p>
</div>
<p>An <a href="https://github.com/turboderp-org/exllamav3">ExLlamaV3</a> quantisation of <a href="https://huggingface.co/cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.0">cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.0</a>, sized for two 24GB cards. The merge itself — what is in it, how the layer curves were built, and how it behaves — is documented on the source repo.</p>
<h2>This quant</h2>
<table>
<tr><th>Format</th><td>EXL3 (ExLlamaV3)</td></tr>
<tr><th>Weights</th><td class="co-num">4.25 bpw</td></tr>
<tr><th>Output head</th><td class="co-num">8-bit (H8)</td></tr>
<tr><th>Weight size on disk</th><td class="co-num">~35 GiB</td></tr>
<tr><th>Architecture</th><td>Llama 3.3 70B, 80 layers</td></tr>
</table>
<p>The 8-bit head is worth the extra space at this bitrate. The output projection is where token probabilities are finally shaped, and quantising it harder than the body tends to show up as vocabulary flattening before it shows up anywhere else.</p>
<h2>Fitting it</h2>
<p>I run this at <strong>40448 tokens of context on a 3090 + 4090 pair</strong> — 48GB of VRAM total. That is close to the ceiling for this configuration; weights alone account for roughly 35 GiB, and the rest is KV cache and overhead.</p>
<p>How much context you actually get depends on your cache quantisation as much as on the model. If you are tight, drop cache precision before dropping context, and expect a single 24GB card not to hold this quant at all.</p>
<h2>Running it</h2>
<p>Llama 3 chat template. The tokenizer is inherited from <a href="https://huggingface.co/allura-org/Llama-3.3-70B-Joyous">Joyous</a> via the source merge, so <code>add_bos_token</code>, <code>pad_token_id</code>, and <code>generation_config.json</code> are all present and correct.</p>
<h3 class="co-h3">Samplers</h3>
<p>What I run, alongside the settings Sapphira's author publishes for Sapphira itself. Treat mine as a starting point rather than a recommendation — taste varies, and a lot of what reads as model character is really sampler character.</p>
<table>
<tr><th></th><th>Sapphira 0.2 (published)</th><th>Mine</th></tr>
<tr><td>Context</td><td class="co-num">—</td><td class="co-num">40448</td></tr>
<tr><td>Temperature</td><td class="co-num">0.95 – 1.1</td><td class="co-num">0.88</td></tr>
<tr><td>Min P</td><td class="co-num">0.015 – 0.03</td><td class="co-num">0.025</td></tr>
<tr><td>Top P</td><td class="co-num">0.97 – 1</td><td class="co-num">1</td></tr>
<tr><td>Typical P</td><td class="co-num">—</td><td class="co-num">1</td></tr>
<tr><td>TFS</td><td class="co-num">—</td><td class="co-num">1</td></tr>
<tr><td>Top K / Top A</td><td class="co-num">—</td><td class="co-num">0</td></tr>
<tr><td>XTC threshold</td><td class="co-num">0.11</td><td class="co-num">off</td></tr>
<tr><td>XTC probability</td><td class="co-num">0.15</td><td class="co-num">off</td></tr>
<tr><td>Repetition penalty</td><td class="co-num">—</td><td class="co-num">1.04</td></tr>
</table>
<p>I sit below Sapphira's temperature range and leave XTC off, with a light repetition penalty doing that work instead. Based on my testing, this merge softens a little logically compared to <a href="https://huggingface.co/cactopus/Omega-Sapphira-L3.3-70B-v1.3">v1.3</a>, and tightening the distribution pulls some of that back. At 4.25bpw I would not push temperature much higher than this — quantisation noise and a loose sampler compound.</p>
<h2>Other quants</h2>
<p>None at present. If you need a different bitrate, the source weights are on the <a href="https://huggingface.co/cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.0">v1.0 repo</a>.</p>
<h2>Before you use it</h2>
<p>All three models underneath this merge are unaligned and intended for adult fiction. It will engage with explicit and violent material without refusing. You are responsible for what you generate. You should be 18 or older.</p>
<p class="co-foot">Quantised with <a href="https://github.com/turboderp-org/exllamav3">ExLlamaV3</a> by Cactopus. Released under the Llama 3.3 Community License. Credit for the underlying models belongs to ReadyArt, BruhzWater, allura-org, and the upstream authors listed on their cards.</p>
</div>