<style>body{font-family:'Quicksand',sans-serif;background:radial-gradient(circle at top,#000000 0%,#100800 85%);color:#fff;margin:0;padding:0;font-size:16px;min-height:100vh}.container{margin:25px;padding:25px;border-radius:16px;background:rgba(5,2,0,.97);border:1px solid rgba(255,140,0,.35);outline:1px solid rgba(255,140,0,.65);outline-offset:-1px;box-shadow:0 0 30px rgba(255,140,0,.25);backdrop-filter:blur(14px);position:relative;overflow:hidden}.container:before{content:'';position:absolute;inset:-2px;border-radius:16px;border:1px solid rgba(255,160,0,.85);pointer-events:none;animation:citrusGlow 2.5s ease-in-out infinite}@keyframes citrusGlow{0%{box-shadow:0 0 8px rgba(255,140,0,.35)}50%{box-shadow:0 0 28px rgba(255,160,0,.85)}100%{box-shadow:0 0 8px rgba(255,140,0,.35)}}.header h1{font-size:32px;color:#ffaa00;margin-bottom:15px;text-shadow:0 0 18px rgba(255,140,0,.95);letter-spacing:1px}h2{font-size:24px;color:#ffaa00;text-shadow:0 0 14px rgba(255,140,0,.75);margin-top:35px;border-bottom:1px solid rgba(255,140,0,.2);padding-bottom:10px}p{color:rgba(255,240,220,.9);line-height:1.7}a{color:#ffaa00;text-decoration:none;transition:.3s ease}a:hover{color:#fff;text-shadow:0 0 12px rgba(255,160,0,.85)}.model-tags{display:flex;flex-wrap:wrap;gap:10px;margin-top:10px}.model-tag{background:rgba(255,140,0,.08);border:1px solid rgba(255,140,0,.25);padding:6px 12px;border-radius:6px;font-size:13px;color:#ffaa00}ul{padding-left:20px}li{margin-bottom:12px;color:rgba(255,240,220,.88)}.card{background:rgba(15,8,0,.92);padding:20px;border-radius:12px;border:1px solid rgba(255,140,0,.25);box-shadow:0 0 20px rgba(255,140,0,.15);margin-top:20px}.special-card{background:linear-gradient(135deg,rgba(30,15,0,0.95),rgba(5,2,0,0.98));border:1px solid rgba(255,180,50,.4);box-shadow:inset 0 0 30px rgba(255,140,0,.1)}.button{display:inline-block;margin:10px 10px 0 0;padding:12px 22px;border-radius:8px;font-weight:600;color:#000;background:linear-gradient(45deg,rgba(255,180,0,.9),rgba(255,120,0,.9));border:1px solid rgba(255,160,0,.5);transition:all .3s ease}.button:hover{transform:translateY(-3px);box-shadow:0 0 20px rgba(255,160,0,.8);color:#fff}.banner{width:100%;border-radius:14px;margin-bottom:20px;border:1px solid rgba(255,140,0,.35);box-shadow:0 0 25px rgba(255,140,0,.25);transition:.3s ease}.banner:hover{transform:scale(1.01);box-shadow:0 0 35px rgba(255,160,0,.45)}code{background:rgba(0,0,0,0.5);padding:2px 6px;border-radius:4px;color:#ffce80;font-family:'Consolas',monospace}pre{background:rgba(0,0,0,0.6);padding:15px;border-radius:8px;border:1px solid rgba(255,140,0,0.2);overflow-x:auto}</style><div class="container"><img class="banner" src="https://huggingface.co/Darkknight535/Void-Citrus-L3.3-70B/resolve/main/ComfyUI_00060_.png" alt="Void-Citrus Banner"><div class="header"><h1>Void-Citrus-L3.3-70B (IQ3_XXS GGUF)</h1></div><div class="card"><p><b>This is the custom GGUF quantization of Void-Citrus-L3.3-70B.</b><br>It is specifically engineered to fit high-performance 70B roleplay into strictly limited VRAM environments (like Dual Tesla T4s or 3090+Offload) without sacrificing the character's voice.</p><div class="model-tags"><span class="model-tag">IQ3_XXS</span><span class="model-tag">Custom Matrix</span><span class="model-tag">BF16 Pipeline</span><span class="model-tag">32GB VRAM Optimized</span></div></div><h2>🍊 The Quantization Difference</h2><div class="card special-card"><p>This is not a standard "click-and-convert" GGUF. It was built using a specialized pipeline to retain maximum coherence at 3-bit compression:</p><ul><li><b>Custom Anime RP Calibration:</b> Unlike standard quants that use generic Wikipedia text (<code>wiki.train.raw</code>) to calculate importance, this model's Importance Matrix (i-mat) was computed using <b>custom Anime Roleplay data</b>. The quantization engine prioritized weights responsible for dialogue, <code>*actions*</code>, and narrative formatting, ensuring the "soul" of the character remains intact.</li><li><b>BF16 Intermediate Source:</b> The conversion bypassed the standard FP16 route. The source model was first converted to <b>BF16 (Brain Float 16)</b> to preserve a higher dynamic range before the final compression step, reducing quantization error.</li><li><b>Surgical Size Fit:</b> The <code>IQ3_XXS</code> format was chosen to land specifically around <b>26.8 GB</b>. This allows the model to fit comfortably on <b>32GB VRAM setups</b> (e.g., Dual T4, Dual 4060 Ti 16GB) with enough room left over for a massive <b>16k+ context window</b> using Q8 KV cache.</li></ul></div><h2>Recommended Settings (Dual GPU)</h2><div class="card"><p>To run this on a <b>Dual 16GB GPU</b> setup (32GB Total) without crashing, use <b>Row Split</b> and <b>Q8 Cache</b>:</p><pre>./llama-cli \
-m void-citrus-l3.3-70b-iq3_xxs-imat.gguf
-p "You are a helpful assistant..."
-n 512
-c 16384
-ngl 99
-sm row
-ctk q8_0
-ctv q8_0</pre><p><i>Note: If on Windows, reduce context to <code>-c 12288</code> to account for WDDM overhead.</i></p></div><h2>Credits</h2><div class="card"><a class="button" href="https://huggingface.co/Darkknight535/Void-Citrus-L3.3-70B" target="_blank">Original Model Page</a></div></div>