Modelpublic

cactopus/Omega-Sapphira-L3.3-70B-v1.3_EXL3_4.25bpw_H8

sourceHugging Facellama3.3updated 22d agoView on Hugging Face

0likes902downloads

Model Card

/ cover art: 3:4, shown whole, edges dissolved into the page / .os-hero { position: relative; padding-top: 26px; } .os-cover { width: 452px; max-width: 90%; margin: 0 auto; position: relative; aspect-ratio: 3 / 4; } .os-cover::before { content: ""; position: absolute; inset: -14% -17% -10%; background: radial-gradient(ellipse 48% 42% at 30% 26%, rgba(214,166,94,0.070), transparent 68%), radial-gradient(ellipse 56% 46% at 62% 64%, rgba(198,150,84,0.050), transparent 72%); filter: blur(28px); z-index: 0; animation: os-breathe 15s ease-in-out infinite; } @keyframes os-breathe { 0%, 100% { opacity: 0.78; transform: scale(1); } 50% { opacity: 1; transform: scale(1.05); } } @media (prefers-reduced-motion: reduce) { .os-cover::before { animation: none; } } .os-cover img { position: relative; z-index: 1; width: 100%; height: 100%; object-fit: contain; display: block; border-radius: 48% 52% 45% 55% / 9% 9% 7% 7%; -webkit-mask-image: radial-gradient(ellipse 86% 88% at 50% 45%, #000 62%, rgba(0,0,0,0.72) 85%, transparent 99%); mask-image: radial-gradient(ellipse 86% 88% at 50% 45%, #000 62%, rgba(0,0,0,0.72) 85%, transparent 99%); } .os-cover::after { content: ""; position: absolute; inset: 0; z-index: 2; pointer-events: none; border-radius: 48% 52% 45% 55% / 9% 9% 7% 7%; background: linear-gradient(108deg, transparent 41%, rgba(255,238,214,0.05) 47%, transparent 53%), radial-gradient(ellipse 82% 86% at 50% 45%, transparent 55%, rgba(20,16,13,0.60) 88%, rgba(20,16,13,0.98) 100%), linear-gradient(to bottom, rgba(20,16,13,0.78) 0%, transparent 10%, transparent 76%, rgba(20,16,13,0.94) 98%); }

.os-head { text-align: center; padding: 26px 0 10px; } .os-name { font-family: Cochin, "Hoefler Text", Baskerville, "Book Antiqua", "Palatino Linotype", Palatino, Georgia, serif; font-size: 2.5em; margin: 0 0 8px; color: #E4CBAB; font-weight: 500; letter-spacing: 0.004em; line-height: 1.12; text-shadow: 0 1px 18px rgba(208,168,104,0.030); } .os-sub { color: #9A8571; font-style: italic; font-size: 1.05em; margin: 0; } .os-pill { margin: 17px 0 0; } .os-pill a { display: inline-block; font-family: monospace; font-size: 0.76em; letter-spacing: 0.13em; text-transform: uppercase; color: #D9A867; border: 1px solid #59432A; background: rgba(200,155,90,0.09); padding: 7px 17px; border-radius: 999px; } .os-pill a:hover { border-color: #D9A867; background: rgba(200,155,90,0.17); }

/ quiet spec rule / .os-spec { display: flex; flex-wrap: wrap; justify-content: center; font-family: monospace; font-size: 0.95em; letter-spacing: 0.08em; text-transform: uppercase; color: #B09C84; border-top: 1px solid #3A2E22; border-bottom: 1px solid #3A2E22; padding: 17px 0; margin: 30px 0 34px; line-height: 1.6; } .os-spec span { padding: 0 26px; border-right: 1px solid #45372A; } .os-spec span:last-child { border-right: none; } .os-spec b { color: #E7CFA6; font-weight: 600; }

/ opening statement / .os-lede { margin: 0 0 30px; padding: 4px 0 4px 24px; border-left: 2px solid #96575C; border-radius: 0 18px 26px 0 / 0 12px 16px 0; } .os-lede p { font-size: 1.15em; line-height: 1.58; color: #E0D3C0; margin: 0 0 12px; } .os-lede p:last-child { margin-bottom: 0; } .os-lede em { color: #D9A867; font-style: italic; }

.os-open::first-letter { float: left; font-size: 3.5em; line-height: 0.82; padding: 6px 12px 0 0; color: #D9A867; font-weight: 600; }

.os h2 { font-size: 1.22em; color: #F2E6D5; font-weight: 600; margin: 46px 0 3px; letter-spacing: -0.01em; } .os h2::before { counter-increment: os-sec; content: counter(os-sec, decimal-leading-zero); font-family: monospace; font-size: 0.6em; font-weight: 400; letter-spacing: 0.1em; color: #C4848A; opacity: 0.85; margin-right: 14px; vertical-align: 0.32em; } .os .os-gloss { color: #8A7460; font-size: 0.78em; text-transform: uppercase; letter-spacing: 0.16em; margin: 0 0 16px; font-family: monospace; padding-left: 34px; } .os h3 { font-size: 1.01em; color: #E8DBC9; font-weight: 600; margin: 28px 0 8px; } .os p, .os li { color: #BFAF9C; } .os strong { color: #E8DBC9; }

.os table { border-collapse: collapse; width: 100%; font-size: 0.92em; margin: 10px 0; } .os th { text-align: left; color: #8A7460; font-weight: 500; font-size: 0.88em; text-transform: uppercase; letter-spacing: 0.09em; font-family: monospace; border-bottom: 1px solid #33281E; padding: 9px 10px 9px 0; } .os td { padding: 9px 10px 9px 0; border-bottom: 1px solid #211A14; color: #BFAF9C; } .os td:first-child, .os th:first-child { color: #E8DBC9; } .os-n { font-variant-numeric: tabular-nums; }

.os pre { background: #0C0907; border: 1px solid #2A211A; border-left: 2px solid #8A6438; border-radius: 0 14px 10px 0 / 0 10px 8px 0; padding: 16px 18px; overflow-x: auto; font-size: 0.82em; line-height: 1.55; box-shadow: inset 0 14px 26px -18px rgba(214,166,96,0.055); } .os code { font-family: "SF Mono", Menlo, Consolas, monospace; color: #C9A87C; } .os a { color: #D9A867; text-decoration: none; border-bottom: 1px solid #59432A; } .os a:hover { border-bottom-color: #D9A867; }

.os-fig { margin: 6px 0 4px; } .os-key { display: flex; gap: 22px; flex-wrap: wrap; font-size: 0.86em; margin: 4px 0 6px; } .os-key span { display: flex; align-items: center; gap: 8px; color: #9A8571; } .os-sw { width: 11px; height: 11px; border-radius: 2px; display: inline-block; } .os-cap { color: #7A6653; font-size: 0.83em; margin: 2px 0 0; }

.os-note { border-left: 2px solid #8A6438; background: rgba(138,100,56,0.07); padding: 15px 19px; margin: 22px 0; border-radius: 0 22px 14px 0 / 0 15px 11px 0; } .os-note p { margin: 0 0 9px; } .os-note p:last-child { margin-bottom: 0; }

.os-import { border: 1px solid #5A4327; border-radius: 0 16px 11px 0 / 0 21px 15px 0; background: radial-gradient(ellipse 58% 120% at 8% 0%, rgba(214,166,96,0.055), transparent 70%), rgba(200,155,90,0.035); padding: 18px 22px; margin: 18px 0 24px; } .os-import p { margin: 0 0 10px; } .os-import p:last-child { margin-bottom: 0; } .os-import-tag { display: block; font-family: monospace; font-size: 0.72em; letter-spacing: 0.17em; text-transform: uppercase; color: #D9A867; margin-bottom: 9px; } .os-import a.os-import-link { font-family: monospace; font-size: 0.95em; word-break: break-all; }

.os-warn { border-left: 2px solid #9E4F38; background: rgba(158,79,56,0.07); padding: 15px 19px; margin: 22px 0; border-radius: 0 14px 24px 0 / 0 10px 16px 0; } .os-warn p { color: #D6B0A0; margin: 0 0 9px; } .os-warn p:last-child { margin-bottom: 0; }

.os-foot { color: #6B5A49; font-size: 0.85em; margin-top: 44px; border-top: 1px solid #241C15; padding-top: 20px; } </style>

<div class="os-head"> <h1 class="os-name">Omega-Sapphira-L3.3-70B-v1.3</h1> <p class="os-sub">EXL3 · 4.25 bits per weight · 8-bit head</p> <p class="os-pill"><a href="https://huggingface.co/cactopus/Omega-Sapphira-L3.3-70B-v1.3">Full weights · bf16</a></p> </div>

<div class="os-lede"> <p>Ten merges came out of the same run. Nine were unremarkable and I have not opened them since. This one I have loaded first, most days, for six months.</p> <p>If you keep one 70B, I would argue for this one.</p> </div>

<p class="os-open">An <a href="https://github.com/turboderp-org/exllamav3">ExLlamaV3</a> quantisation of <a href="https://huggingface.co/cactopus/Omega-Sapphira-L3.3-70B-v1.3">cactopus/Omega-Sapphira-L3.3-70B-v1.3</a>, cut to fit two 24GB cards. Omega Directive v2.1 for structure, Sapphira 0.2 for prose, blended with separate depth-dependent curves for attention and feed-forward blocks rather than one flat ratio. The full breakdown lives on the source repo.</p>

<h2>The culture</h2> <p class="os-gloss">this quant</p>

<table> <tr><th>Format</th><td>EXL3 (ExLlamaV3)</td></tr> <tr><th>Weights</th><td class="os-n">4.25 bpw</td></tr> <tr><th>Output head</th><td class="os-n">8-bit (H8)</td></tr> <tr><th>Weight size on disk</th><td class="os-n">~35 GiB</td></tr> <tr><th>Architecture</th><td>Llama 3.3 70B, 80 layers</td></tr> </table>

<p>The 8-bit head earns its space at this bitrate. The output projection is where token probabilities are finally shaped, and quantising it harder than the body shows up as vocabulary flattening before it shows up anywhere else. On a merge chosen partly for its prose, that is the wrong economy.</p>

<h2>Vessel requirements</h2> <p class="os-gloss">fitting it</p>

<p><strong>40448 tokens of context on a 3090 + 4090 pair</strong> — 48GB of VRAM total. That is near the ceiling: weights alone account for roughly 35 GiB, and the remainder goes to KV cache and overhead.</p>

<p>How much context you actually get depends on cache quantisation as much as on the model. If you are tight, drop cache precision before dropping context. A single 24GB card will not hold this.</p>

<h2>Culture conditions</h2> <p class="os-gloss">running it</p>

<p>Llama 3 chat template.</p>

<div class="os-note"> <p>The tokenizer comes from the Omega base. Those files use list-format BPE merges and omit <code>addbostoken</code>, <code>padtokenid</code> and <code>generationconfig.json</code>.</p> <p>Vocabulary and merge rules are stock Llama 3.3, so nothing about the weights depends on this. But if your loader does not prepend <code><|beginof_text|></code> by default, set it explicitly — a missing BOS shifts every position and degrades output quietly rather than obviously.</p> </div>

<h3>Samplers</h3>

<div class="os-import"> <span class="os-import-tag">One-click setup · SillyTavern</span> <p>Rather than set twenty fields by hand, import my whole configuration — instruct template, context template, system prompt and sampler preset:</p> <p><a class="os-import-link" href="https://pastebin.com/nctn66zg">https://pastebin.com/nctn66zg</a></p> <p>Save the text and rename it to end in <code>.json</code>. In SillyTavern, click the <strong>A</strong> button (AI Response Formatting) in the top bar, then <strong>Master Import</strong> at the top right of that panel.</p> <p>It also sets <code>addbostoken: true</code>, which matters here — see the tokenizer note above. The preset carries <code>max_length: 40448</code>, which is the 48GB figure above — drop it to whatever your own setup fits.</p> </div>

<p>If you would rather set them by hand: known-good rather than required. This one holds together across a fair band, so treat the numbers as a place to start rather than a target to hit:</p>

<pre><code>Temperature: 0.9 - 1.1 Min P: 0.02 - 0.025 Top P: 0.96 Top K: 90 TFS: 1 (off) Typical P: 1 Repetition Penalty: 1.02</code></pre>

<p><strong>Mirostat is worth experimenting with.</strong> The preset ships with it off (mode 0), but I often prefer mode 2 — tau 6 and eta 0.1 are already in there, so it is a single dropdown change. Treat it as either/or rather than additive: with Mirostat on, most backends hand it the tail and largely bypass Top-P, Top-K and Min-P.</p>

<p>One thing worth knowing if you also run <a href="https://huggingface.co/cactopus/OmegaSapphiraJoyous-L3.3-70B-v1.1">the Joyous graft</a>: it wants the opposite shape, TFS on and Top-P off. Presets do not travel between the two.</p>

<p>Sapphira's card also names a Deep Cogito instruction template alongside Llama 3, inherited from the cogito base underneath it. Worth trying if you want behaviour closer to Sapphira's own.</p>

<h2>Related cultures</h2> <p class="os-gloss">elsewhere in the family</p>

<p>The mommy workhorse of the line. Every daughter merge I have released is grafted onto this one, and not one of them has inherited its manners.</p>

<p><a href="https://huggingface.co/cactopus/Omega-Sapphira-L3.3-70B-v1.3">Source weights</a> — full bf16, with the merge configuration and the layer-depth breakdown.</p>

<p><a href="https://huggingface.co/cactopus/OmegaSapphiraJoyous-L3.3-70B-v1.1">Omega-Sapphira-Joyous v1.1</a> — a thin Joyous graft over this one. Different prose rather than better, and a considerably more stubborn daughter: it wants specific settings before it will behave. Also <a href="https://huggingface.co/cactopus/OmegaSapphiraJoyous-L3.3-70B-v1.1EXL34.25bpw_H8">at this bitrate</a>.</p>

<h2>Handling</h2> <p class="os-gloss">before you use it</p>

<div class="os-warn"> <p>Both models underneath this merge are unaligned and intended for adult fiction. It will engage with explicit and violent material without refusing.</p> <p>You are responsible for what you generate. You should be 18 or older.</p> </div>

<p class="os-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, and the upstream authors listed on their cards.</p>

</div> </div>