Instructions to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp
# Run inference directly in the terminal:
llama cli -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp
# Run inference directly in the terminal:
llama cli -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp
# Run inference directly in the terminal:
./build/bin/llama cli -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Use Docker

docker model run hf.co/Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Llama new
LM Studio
Jan

vLLM

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Ollama
How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with Ollama:
```
ollama run hf.co/Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp
```
Unsloth Desktop

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with Docker Model Runner:
```
docker model run hf.co/Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp
```

Lemonade

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Run and chat with the model

lemonade run user.Split-Untied-31B-MaxOptimaV2-ik_llama.cpp-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Austriani/Split-Untied-31B-MaxOptimaV2-ik_llama.cpp" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Split-Untied-31B-MaxOptimaV2-ik_llama.cpp / README.md

Austriani

Update README.md

adfc8fe verified 8 days ago

preview code

Raw History Blame Contribute Delete

14.9 kB

---
license: apache-2.0
license_link: https://huggingface.co/Blazed-Forge/Split-Untied-31B/blob/main/LICENSE
base_model: Blazed-Forge/Split-Untied-31B
pipeline_tag: text-generation
library_name: ik_llama.cpp
language:
- en
tags:
  - gemma-4
  - gemma
  - split-untied
  - 31b
  - gguf
  - quantized
  - quantization
  - ik-llama.cpp
  - iqk
  - trellis-quant
  - maxoptima
  - imatrix
  - v3-calibration
  - uncensored-imatrix
  - roleplay
  - rp
  - creative-writing
  - storywriting
  - fiction
  - dark-themes
  - uncensored
  - chat
  - conversational
  - text-generation
  - merge
  - austriani
  - nimbz
  - blazed-forge
---
<style>
.mx-universe {
  max-width: 960px;
  margin: 0 auto;
  padding: 40px 20px;
  background-color: #0f172a;
  background-image: 
    radial-gradient(circle at 15% 20%, rgba(226, 184, 107, 0.05) 0%, transparent 40%),
    radial-gradient(circle at 85% 80%, rgba(168, 85, 247, 0.06) 0%, transparent 50%),
    radial-gradient(circle at 50% 50%, rgba(168, 85, 247, 0.03) 0%, transparent 60%);
  position: relative;
  border-radius: 16px;
  overflow: hidden;
  font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
}
.mx-universe::before {
  content: '';
  position: absolute;
  inset: 0;
  background-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 200 200' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='noiseFilter'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='0.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23noiseFilter)'/%3E%3C/svg%3E");
  opacity: 0.03;
  pointer-events: none;
  z-index: 0;
}

.mx-glass {
  background: rgba(15, 23, 42, 0.75);
  backdrop-filter: blur(20px);
  -webkit-backdrop-filter: blur(20px);
  border: 1px solid rgba(226, 184, 107, 0.15);
  border-top: 3px solid transparent;
  border-image: linear-gradient(90deg, #78350f, #e2b86b, #a855f7, #e2b86b, #78350f) 1;
  box-shadow: 
    0 30px 60px rgba(0, 0, 0, 0.6),
    inset 0 1px 0 rgba(255, 255, 255, 0.05),
    0 0 40px rgba(168, 85, 247, 0.08);
  padding: 60px;
  border-radius: 0 0 16px 16px;
  position: relative;
  z-index: 2;
}
.mx-header {
  text-align: center;
  margin-bottom: 50px;
  position: relative;
}

.mx-kicker {
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.8rem;
  font-weight: 700;
  letter-spacing: 5px;
  text-transform: uppercase;
  color: #e2b86b;
  margin-bottom: 20px;
  text-shadow: 0 0 15px rgba(226, 184, 107, 0.3);
}

.mx-title {
  font-family: "Times New Roman", Times, serif;
  font-size: 3.2rem;
  font-weight: 700;
  line-height: 1.1;
  margin: 0 0 20px;
  background: linear-gradient(135deg, #f8fafc 0%, #cbd5e1 50%, #64748b 100%);
  -webkit-background-clip: text;
  -webkit-text-fill-color: transparent;
  filter: drop-shadow(0 4px 10px rgba(0,0,0,0.5));
}

.mx-subtitle {
  font-size: 1.15rem;
  color: #94a3b8;
  font-weight: 300;
  letter-spacing: 1px;
}

.mx-glow-line {
  width: 200px;
  height: 2px;
  background: linear-gradient(90deg, transparent, #a855f7, #e2b86b, #a855f7, transparent);
  margin: 30px auto 0;
  box-shadow: 0 0 20px rgba(168, 85, 247, 0.6);
}

.mx-grid {
  display: grid;
  grid-template-columns: repeat(4, 1fr);
  gap: 20px;
  margin-bottom: 50px;
}

.mx-stat {
  background: rgba(30, 41, 59, 0.6);
  border: 1px solid rgba(255, 255, 255, 0.05);
  border-radius: 12px;
  padding: 30px 20px;
  text-align: center;
  transition: all 0.4s cubic-bezier(0.175, 0.885, 0.32, 1.275);
  box-shadow: 0 10px 20px rgba(0,0,0,0.2);
}

.mx-stat:hover {
  transform: translateY(-5px) scale(1.02);
  border-color: rgba(168, 85, 247, 0.4);
  box-shadow: 0 15px 30px rgba(0,0,0,0.4), 0 0 20px rgba(168, 85, 247, 0.15);
}

.mx-stat-val {
  font-family: "Times New Roman", Times, serif;
  font-size: 2.4rem;
  font-weight: 700;
  color: #e2b86b;
  line-height: 1;
  margin-bottom: 12px;
  text-shadow: 0 0 20px rgba(226, 184, 107, 0.2);
}

.mx-stat-lbl {
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.7rem;
  font-weight: 700;
  letter-spacing: 2px;
  text-transform: uppercase;
  color: #64748b;
}

.mx-text {
  font-size: 1.05rem;
  color: #cbd5e1;
  line-height: 1.9;
  margin-bottom: 25px;
}

.mx-text strong {
  color: #f8fafc;
  font-weight: 600;
}

.mx-text a {
  color: #e2b86b;
  text-decoration: none;
  border-bottom: 1px dashed #e2b86b;
  transition: all 0.3s;
}

.mx-text a:hover {
  color: #fef3c7;
  border-bottom-style: solid;
  text-shadow: 0 0 10px rgba(226, 184, 107, 0.4);
}

.mx-h2 {
  font-family: "Times New Roman", Times, serif;
  font-size: 1.5rem;
  font-weight: 700;
  color: #f8fafc;
  margin: 60px 0 25px;
  display: flex;
  align-items: center;
  gap: 15px;
  letter-spacing: 1px;
}

.mx-h2::before {
  content: '';
  display: block;
  width: 10px;
  height: 10px;
  background: #a855f7;
  border-radius: 50%;
  box-shadow: 0 0 15px #a855f7;
}

.mx-h2::after {
  content: '';
  flex: 1;
  height: 1px;
  background: linear-gradient(90deg, rgba(168, 85, 247, 0.4), transparent);
}

.mx-table-wrap {
  background: rgba(15, 23, 42, 0.8);
  border: 1px solid rgba(255, 255, 255, 0.05);
  border-radius: 12px;
  overflow: hidden;
  margin-bottom: 30px;
  box-shadow: 0 10px 30px rgba(0,0,0,0.3);
}

.mx-table {
  width: 100%;
  border-collapse: collapse;
}

.mx-table th {
  background: rgba(30, 41, 59, 0.9);
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.75rem;
  font-weight: 700;
  letter-spacing: 2px;
  text-transform: uppercase;
  color: #e2b86b;
  padding: 20px;
  text-align: left;
  border-bottom: 2px solid rgba(226, 184, 107, 0.2);
}

.mx-table td {
  padding: 18px 20px;
  font-size: 0.95rem;
  color: #94a3b8;
  border-bottom: 1px solid rgba(255, 255, 255, 0.03);
  transition: background 0.3s;
}

.mx-table tr:hover td {
  background: rgba(168, 85, 247, 0.04);
  color: #e2e8f0;
}

.mx-table tr:last-child td {
  border-bottom: none;
}

.mx-table td:first-child {
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-weight: 700;
  color: #e2e8f0;
  white-space: nowrap;
}

.mx-code {
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.85em;
  background: rgba(226, 184, 107, 0.08);
  padding: 3px 8px;
  border-radius: 4px;
  color: #e2b86b;
  border: 1px solid rgba(226, 184, 107, 0.15);
}

.mx-dataset-bar {
  display: flex;
  height: 24px;
  border-radius: 6px;
  overflow: hidden;
  margin: 20px 0 10px;
  box-shadow: 0 4px 10px rgba(0,0,0,0.3);
}

.mx-dataset-segment {
  display: flex;
  align-items: center;
  justify-content: center;
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.65rem;
  font-weight: 700;
  color: #0f172a;
  transition: width 0.5s ease;
}

.mx-seg-1 { background: #e2b86b; width: 40%; }
.mx-seg-2 { background: #cbd5e1; width: 25%; }
.mx-seg-3 { background: #94a3b8; width: 20%; }
.mx-seg-4 { background: #64748b; width: 15%; }

.mx-dataset-legend {
  display: flex;
  flex-wrap: wrap;
  gap: 15px;
  margin-bottom: 30px;
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.75rem;
  color: #94a3b8;
}

.mx-legend-item {
  display: flex;
  align-items: center;
  gap: 6px;
}

.mx-legend-dot {
  width: 10px;
  height: 10px;
  border-radius: 50%;
}

.mx-callout {
  background: linear-gradient(135deg, rgba(226, 184, 107, 0.04), rgba(168, 85, 247, 0.04));
  border: 1px solid rgba(168, 85, 247, 0.2);
  border-left: 4px solid #a855f7;
  border-radius: 8px;
  padding: 30px;
  margin: 40px 0;
  position: relative;
  overflow: hidden;
  box-shadow: inset 0 0 30px rgba(0,0,0,0.4);
}

.mx-callout p {
  margin: 0;
  color: #e2e8f0;
  font-size: 1rem;
  line-height: 1.8;
}

.mx-callout strong {
  color: #e2b86b;
}

.mx-footer {
  margin-top: 60px;
  padding-top: 30px;
  border-top: 1px solid rgba(255, 255, 255, 0.05);
  display: flex;
  justify-content: space-between;
  flex-wrap: wrap;
  gap: 20px;
  font-family: "JetBrains Mono", "Courier New", monospace;
  font-size: 0.75rem;
  color: #475569;
  text-transform: uppercase;
  letter-spacing: 2px;
}

@media (max-width: 800px) {
  .mx-glass { padding: 30px 20px; }
  .mx-grid { grid-template-columns: 1fr 1fr; }
  .mx-title { font-size: 2.2rem; }
  .mx-dataset-bar { height: auto; flex-direction: column; }
  .mx-dataset-segment { width: 100% !important; height: 30px; }
}

@media (max-width: 500px) {
  .mx-grid { grid-template-columns: 1fr; }
}
</style>

<div class="mx-universe">
<div class="mx-glass">

<div class="mx-header">
<div class="mx-kicker">Bespoke Quantization // ik_llama.cpp</div>
<h1 class="mx-title">Split-Untied-31B<br>MaxOptimaV2</h1>
<div class="mx-subtitle">A meticulously crafted IQK mix for the darker side of roleplay.</div>
<div class="mx-glow-line"></div>
</div>
<div class="mx-grid">
<div class="mx-stat">
<div class="mx-stat-val">150</div>
<div class="mx-stat-lbl">Cycles</div>
</div>
<div class="mx-stat">
<div class="mx-stat-val">300</div>
<div class="mx-stat-lbl">Chunks</div>
</div>
<div class="mx-stat">
<div class="mx-stat-val">1024</div>
<div class="mx-stat-lbl">Context</div>
</div>
<div class="mx-stat">
<div class="mx-stat-val">4h+</div>
<div class="mx-stat-lbl">Compute Time</div>
</div>
</div>
<div class="mx-text">
This repository hosts a highly specialized <strong>ik_llama.cpp quantization</strong> of <a href="https://huggingface.co/Blazed-Forge/Split-Untied-31B">Blazed-Forge/Split-Untied-31B</a>. The base model is Nimbz's 31B Gemma 4 merge, built through a three-phase StyleSwap process that prioritizes <strong>creative writing, darker themes, and expressive freedom</strong>. It excels at imaginative, less constrained narratives and deep character roleplay.
</div>

<div class="mx-text">
Standard quantization methods treat every layer with the same broad strokes. <strong>MaxOptimaV2</strong> is different. This repository is dedicated entirely to this bespoke, layer-by-layer recipe. It was designed to compete directly with standard <span class="mx-code">IQ4_KS</span>, offering quality on par with or slightly better than a generic <span class="mx-code">Q4_K_M</span> while maintaining a smaller file size. Every tensor was deliberately assigned to protect the model's creative voice.
</div>
<div class="mx-h2">The MaxOptimaV2 Strategy</div>
<div class="mx-text">
To preserve the expressive nature of Split-Untied, standard uniform 4-bit compression is insufficient. The MaxOptimaV2 script uses a surgical sandwich approach. It safeguards the most critical attention and feed-forward layers at <span class="mx-code">IQ5_K</span> and <span class="mx-code">Q6_K</span>, while compressing the bulk of the model to <span class="mx-code">IQ4_KSS</span> and <span class="mx-code">IQ4_KS</span>.
</div>
<div class="mx-table-wrap">
<table class="mx-table">
<thead>
<tr><th>Layer Group</th><th>Quant</th><th>Rationale</th></tr>
</thead>
<tbody>
<tr><td>output.weight</td><td>Q6_K</td><td>Final projection kept at high precision to preserve vocabulary richness.</td></tr>
<tr><td>token_embd.weight</td><td>Q5_K</td><td>Embeddings balanced for size versus semantic depth.</td></tr>
<tr><td>blk.0 / blk.58 attn_v</td><td>IQ5_K</td><td>Edge attention layers protected at 5-bit.</td></tr>
<tr><td>blk.5, 11, 17, 23, 29, 35, 41, 47, 53, 59 attn_k</td><td>IQ5_K</td><td>Key attention layers kept sharp for multi-turn context tracking.</td></tr>
<tr><td>blk.0 / blk.59 ffn_down</td><td>IQ5_K</td><td>Boundary FFN layers preserved.</td></tr>
<tr><td>All other ffn_down</td><td>IQ4_KSS</td><td>Balanced 4-bit for the bulk of the feed-forward network.</td></tr>
<tr><td>blk.50–59 ffn_gate / ffn_up</td><td>IQ4_KSS</td><td>Late FFN layers kept at 4-bit to protect output style.</td></tr>
<tr><td>blk.0–19 ffn_gate / ffn_up</td><td>IQ4_KSS</td><td>Early FFN layers similarly balanced.</td></tr>
<tr><td>All remaining tensors</td><td>IQ4_KS</td><td>Default fallback for maximum compression at 4-bit.</td></tr>
</tbody>
</table>
</div>
<div class="mx-h2">The V3 Calibration Dataset</div>
<div class="mx-text">
The importance matrix for this quantization was not generated lightly. It required over four hours of dedicated compute time, processing 300 chunks of domain-specific data at a 1024-token context window across 150 cycles. 
</div>
<div class="mx-text">
Crucially, this imatrix was built using a custom <strong>V3 Calibration Dataset</strong>. We took the <a href="https://huggingface.co/datasets/OpenIntelligenceNet/Uncensored-Imatrix-Calibration-Mixed-Data">OpenIntelligenceNet Mixed Data</a> and completely stripped out all Polyglot, Chain-of-Thought, and Factual knowledge data. What remains is a laser-focused dataset containing <strong>only Uncensored and Roleplay data</strong>. This ensures the quantization weights are perfectly optimized for creative and unfiltered inference, rather than wasting precision on general factual recall.
</div>
<div class="mx-dataset-bar">
<div class="mx-dataset-segment mx-seg-1">40%</div>
<div class="mx-dataset-segment mx-seg-2">25%</div>
<div class="mx-dataset-segment mx-seg-3">20%</div>
<div class="mx-dataset-segment mx-seg-4">15%</div>
</div>
<div class="mx-dataset-legend">
<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-1"></div> 40% v2_stripped</div>
<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-2"></div> 25% DPO Roleplay (ath)</div>
<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-3"></div> 20% Character Cards (gry)</div>
<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-4"></div> 15% SynthRP (epi)</div>
</div>

<div class="mx-text" style="font-size: 0.9rem; color: #64748b;">
The v2_stripped source consists of the Uncensored Alpaca Blend and Multi-Turn Roleplay &amp; Fiction. The ath source provides DPO Pairs for NSFW roleplay. The gry source is Sonnet 3.5 character card roleplay. The epi source is synthetic roleplay generations.
</div>
<div class="mx-callout">
<p><strong>Competitive Positioning:</strong> MaxOptimaV2 targets the same niche as <span class="mx-code">IQ4_KS</span> but with a finer-grained weight distribution. In practice, it offers comparable or slightly better roleplay coherence than <span class="mx-code">Q4_K_M</span> while producing a smaller file. It is not a universal upgrade. It is a purpose-built tool for creative and uncensored inference.</p>
</div>

<div class="mx-h2">Compatibility</div>
<div class="mx-text">
<strong>MaxOptimaV2 is exclusive to ik_llama.cpp.</strong> It will not load in mainline llama.cpp. If you require mainline compatibility or standard formats like BF16, please refer to my previous repository: <a href="https://huggingface.co/Austriani/Split-Untied-31B-ik_llama.cpp">Austriani/Split-Untied-31B-ik_llama.cpp</a>.
</div>

<div class="mx-footer">
<div>Austriani // Split-Untied-31B MaxOptimaV2</div>
<div>Base: Blazed-Forge/Split-Untied-31B // V3 Calibration</div>
</div>

</div>
</div>

	---
	license: apache-2.0
	license_link: https://huggingface.co/Blazed-Forge/Split-Untied-31B/blob/main/LICENSE
	base_model: Blazed-Forge/Split-Untied-31B
	pipeline_tag: text-generation
	library_name: ik_llama.cpp
	language:
	- en
	tags:
	- gemma-4
	- gemma
	- split-untied
	- 31b
	- gguf
	- quantized
	- quantization
	- ik-llama.cpp
	- iqk
	- trellis-quant
	- maxoptima
	- imatrix
	- v3-calibration
	- uncensored-imatrix
	- roleplay
	- rp
	- creative-writing
	- storywriting
	- fiction
	- dark-themes
	- uncensored
	- chat
	- conversational
	- text-generation
	- merge
	- austriani
	- nimbz
	- blazed-forge
	---

	<style>
	.mx-universe {
	max-width: 960px;
	margin: 0 auto;
	padding: 40px 20px;
	background-color: #0f172a;
	background-image:
	radial-gradient(circle at 15% 20%, rgba(226, 184, 107, 0.05) 0%, transparent 40%),
	radial-gradient(circle at 85% 80%, rgba(168, 85, 247, 0.06) 0%, transparent 50%),
	radial-gradient(circle at 50% 50%, rgba(168, 85, 247, 0.03) 0%, transparent 60%);
	position: relative;
	border-radius: 16px;
	overflow: hidden;
	font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
	}

	.mx-universe::before {
	content: '';
	position: absolute;
	inset: 0;
	background-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 200 200' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='noiseFilter'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='0.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23noiseFilter)'/%3E%3C/svg%3E");
	opacity: 0.03;
	pointer-events: none;
	z-index: 0;
	}

	.mx-glass {
	background: rgba(15, 23, 42, 0.75);
	backdrop-filter: blur(20px);
	-webkit-backdrop-filter: blur(20px);
	border: 1px solid rgba(226, 184, 107, 0.15);
	border-top: 3px solid transparent;
	border-image: linear-gradient(90deg, #78350f, #e2b86b, #a855f7, #e2b86b, #78350f) 1;
	box-shadow:
	0 30px 60px rgba(0, 0, 0, 0.6),
	inset 0 1px 0 rgba(255, 255, 255, 0.05),
	0 0 40px rgba(168, 85, 247, 0.08);
	padding: 60px;
	border-radius: 0 0 16px 16px;
	position: relative;
	z-index: 2;
	}

	.mx-header {
	text-align: center;
	margin-bottom: 50px;
	position: relative;
	}

	.mx-kicker {
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.8rem;
	font-weight: 700;
	letter-spacing: 5px;
	text-transform: uppercase;
	color: #e2b86b;
	margin-bottom: 20px;
	text-shadow: 0 0 15px rgba(226, 184, 107, 0.3);
	}

	.mx-title {
	font-family: "Times New Roman", Times, serif;
	font-size: 3.2rem;
	font-weight: 700;
	line-height: 1.1;
	margin: 0 0 20px;
	background: linear-gradient(135deg, #f8fafc 0%, #cbd5e1 50%, #64748b 100%);
	-webkit-background-clip: text;
	-webkit-text-fill-color: transparent;
	filter: drop-shadow(0 4px 10px rgba(0,0,0,0.5));
	}

	.mx-subtitle {
	font-size: 1.15rem;
	color: #94a3b8;
	font-weight: 300;
	letter-spacing: 1px;
	}

	.mx-glow-line {
	width: 200px;
	height: 2px;
	background: linear-gradient(90deg, transparent, #a855f7, #e2b86b, #a855f7, transparent);
	margin: 30px auto 0;
	box-shadow: 0 0 20px rgba(168, 85, 247, 0.6);
	}

	.mx-grid {
	display: grid;
	grid-template-columns: repeat(4, 1fr);
	gap: 20px;
	margin-bottom: 50px;
	}

	.mx-stat {
	background: rgba(30, 41, 59, 0.6);
	border: 1px solid rgba(255, 255, 255, 0.05);
	border-radius: 12px;
	padding: 30px 20px;
	text-align: center;
	transition: all 0.4s cubic-bezier(0.175, 0.885, 0.32, 1.275);
	box-shadow: 0 10px 20px rgba(0,0,0,0.2);
	}

	.mx-stat:hover {
	transform: translateY(-5px) scale(1.02);
	border-color: rgba(168, 85, 247, 0.4);
	box-shadow: 0 15px 30px rgba(0,0,0,0.4), 0 0 20px rgba(168, 85, 247, 0.15);
	}

	.mx-stat-val {
	font-family: "Times New Roman", Times, serif;
	font-size: 2.4rem;
	font-weight: 700;
	color: #e2b86b;
	line-height: 1;
	margin-bottom: 12px;
	text-shadow: 0 0 20px rgba(226, 184, 107, 0.2);
	}

	.mx-stat-lbl {
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.7rem;
	font-weight: 700;
	letter-spacing: 2px;
	text-transform: uppercase;
	color: #64748b;
	}

	.mx-text {
	font-size: 1.05rem;
	color: #cbd5e1;
	line-height: 1.9;
	margin-bottom: 25px;
	}

	.mx-text strong {
	color: #f8fafc;
	font-weight: 600;
	}

	.mx-text a {
	color: #e2b86b;
	text-decoration: none;
	border-bottom: 1px dashed #e2b86b;
	transition: all 0.3s;
	}

	.mx-text a:hover {
	color: #fef3c7;
	border-bottom-style: solid;
	text-shadow: 0 0 10px rgba(226, 184, 107, 0.4);
	}

	.mx-h2 {
	font-family: "Times New Roman", Times, serif;
	font-size: 1.5rem;
	font-weight: 700;
	color: #f8fafc;
	margin: 60px 0 25px;
	display: flex;
	align-items: center;
	gap: 15px;
	letter-spacing: 1px;
	}

	.mx-h2::before {
	content: '';
	display: block;
	width: 10px;
	height: 10px;
	background: #a855f7;
	border-radius: 50%;
	box-shadow: 0 0 15px #a855f7;
	}

	.mx-h2::after {
	content: '';
	flex: 1;
	height: 1px;
	background: linear-gradient(90deg, rgba(168, 85, 247, 0.4), transparent);
	}

	.mx-table-wrap {
	background: rgba(15, 23, 42, 0.8);
	border: 1px solid rgba(255, 255, 255, 0.05);
	border-radius: 12px;
	overflow: hidden;
	margin-bottom: 30px;
	box-shadow: 0 10px 30px rgba(0,0,0,0.3);
	}

	.mx-table {
	width: 100%;
	border-collapse: collapse;
	}

	.mx-table th {
	background: rgba(30, 41, 59, 0.9);
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.75rem;
	font-weight: 700;
	letter-spacing: 2px;
	text-transform: uppercase;
	color: #e2b86b;
	padding: 20px;
	text-align: left;
	border-bottom: 2px solid rgba(226, 184, 107, 0.2);
	}

	.mx-table td {
	padding: 18px 20px;
	font-size: 0.95rem;
	color: #94a3b8;
	border-bottom: 1px solid rgba(255, 255, 255, 0.03);
	transition: background 0.3s;
	}

	.mx-table tr:hover td {
	background: rgba(168, 85, 247, 0.04);
	color: #e2e8f0;
	}

	.mx-table tr:last-child td {
	border-bottom: none;
	}

	.mx-table td:first-child {
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-weight: 700;
	color: #e2e8f0;
	white-space: nowrap;
	}

	.mx-code {
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.85em;
	background: rgba(226, 184, 107, 0.08);
	padding: 3px 8px;
	border-radius: 4px;
	color: #e2b86b;
	border: 1px solid rgba(226, 184, 107, 0.15);
	}

	.mx-dataset-bar {
	display: flex;
	height: 24px;
	border-radius: 6px;
	overflow: hidden;
	margin: 20px 0 10px;
	box-shadow: 0 4px 10px rgba(0,0,0,0.3);
	}

	.mx-dataset-segment {
	display: flex;
	align-items: center;
	justify-content: center;
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.65rem;
	font-weight: 700;
	color: #0f172a;
	transition: width 0.5s ease;
	}

	.mx-seg-1 { background: #e2b86b; width: 40%; }
	.mx-seg-2 { background: #cbd5e1; width: 25%; }
	.mx-seg-3 { background: #94a3b8; width: 20%; }
	.mx-seg-4 { background: #64748b; width: 15%; }

	.mx-dataset-legend {
	display: flex;
	flex-wrap: wrap;
	gap: 15px;
	margin-bottom: 30px;
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.75rem;
	color: #94a3b8;
	}

	.mx-legend-item {
	display: flex;
	align-items: center;
	gap: 6px;
	}

	.mx-legend-dot {
	width: 10px;
	height: 10px;
	border-radius: 50%;
	}

	.mx-callout {
	background: linear-gradient(135deg, rgba(226, 184, 107, 0.04), rgba(168, 85, 247, 0.04));
	border: 1px solid rgba(168, 85, 247, 0.2);
	border-left: 4px solid #a855f7;
	border-radius: 8px;
	padding: 30px;
	margin: 40px 0;
	position: relative;
	overflow: hidden;
	box-shadow: inset 0 0 30px rgba(0,0,0,0.4);
	}

	.mx-callout p {
	margin: 0;
	color: #e2e8f0;
	font-size: 1rem;
	line-height: 1.8;
	}

	.mx-callout strong {
	color: #e2b86b;
	}

	.mx-footer {
	margin-top: 60px;
	padding-top: 30px;
	border-top: 1px solid rgba(255, 255, 255, 0.05);
	display: flex;
	justify-content: space-between;
	flex-wrap: wrap;
	gap: 20px;
	font-family: "JetBrains Mono", "Courier New", monospace;
	font-size: 0.75rem;
	color: #475569;
	text-transform: uppercase;
	letter-spacing: 2px;
	}

	@media (max-width: 800px) {
	.mx-glass { padding: 30px 20px; }
	.mx-grid { grid-template-columns: 1fr 1fr; }
	.mx-title { font-size: 2.2rem; }
	.mx-dataset-bar { height: auto; flex-direction: column; }
	.mx-dataset-segment { width: 100% !important; height: 30px; }
	}

	@media (max-width: 500px) {
	.mx-grid { grid-template-columns: 1fr; }
	}
	</style>

	<div class="mx-universe">
	<div class="mx-glass">

	<div class="mx-header">
	<div class="mx-kicker">Bespoke Quantization // ik_llama.cpp</div>
	<h1 class="mx-title">Split-Untied-31B<br>MaxOptimaV2</h1>
	<div class="mx-subtitle">A meticulously crafted IQK mix for the darker side of roleplay.</div>
	<div class="mx-glow-line"></div>
	</div>

	<div class="mx-grid">
	<div class="mx-stat">
	<div class="mx-stat-val">150</div>
	<div class="mx-stat-lbl">Cycles</div>
	</div>
	<div class="mx-stat">
	<div class="mx-stat-val">300</div>
	<div class="mx-stat-lbl">Chunks</div>
	</div>
	<div class="mx-stat">
	<div class="mx-stat-val">1024</div>
	<div class="mx-stat-lbl">Context</div>
	</div>
	<div class="mx-stat">
	<div class="mx-stat-val">4h+</div>
	<div class="mx-stat-lbl">Compute Time</div>
	</div>
	</div>

	<div class="mx-text">
	This repository hosts a highly specialized <strong>ik_llama.cpp quantization</strong> of <a href="https://huggingface.co/Blazed-Forge/Split-Untied-31B">Blazed-Forge/Split-Untied-31B</a>. The base model is Nimbz's 31B Gemma 4 merge, built through a three-phase StyleSwap process that prioritizes <strong>creative writing, darker themes, and expressive freedom</strong>. It excels at imaginative, less constrained narratives and deep character roleplay.
	</div>

	<div class="mx-text">
	Standard quantization methods treat every layer with the same broad strokes. <strong>MaxOptimaV2</strong> is different. This repository is dedicated entirely to this bespoke, layer-by-layer recipe. It was designed to compete directly with standard <span class="mx-code">IQ4_KS</span>, offering quality on par with or slightly better than a generic <span class="mx-code">Q4_K_M</span> while maintaining a smaller file size. Every tensor was deliberately assigned to protect the model's creative voice.
	</div>

	<div class="mx-h2">The MaxOptimaV2 Strategy</div>

	<div class="mx-text">
	To preserve the expressive nature of Split-Untied, standard uniform 4-bit compression is insufficient. The MaxOptimaV2 script uses a surgical sandwich approach. It safeguards the most critical attention and feed-forward layers at <span class="mx-code">IQ5_K</span> and <span class="mx-code">Q6_K</span>, while compressing the bulk of the model to <span class="mx-code">IQ4_KSS</span> and <span class="mx-code">IQ4_KS</span>.
	</div>

	<div class="mx-table-wrap">
	<table class="mx-table">
	<thead>
	<tr><th>Layer Group</th><th>Quant</th><th>Rationale</th></tr>
	</thead>
	<tbody>
	<tr><td>output.weight</td><td>Q6_K</td><td>Final projection kept at high precision to preserve vocabulary richness.</td></tr>
	<tr><td>token_embd.weight</td><td>Q5_K</td><td>Embeddings balanced for size versus semantic depth.</td></tr>
	<tr><td>blk.0 / blk.58 attn_v</td><td>IQ5_K</td><td>Edge attention layers protected at 5-bit.</td></tr>
	<tr><td>blk.5, 11, 17, 23, 29, 35, 41, 47, 53, 59 attn_k</td><td>IQ5_K</td><td>Key attention layers kept sharp for multi-turn context tracking.</td></tr>
	<tr><td>blk.0 / blk.59 ffn_down</td><td>IQ5_K</td><td>Boundary FFN layers preserved.</td></tr>
	<tr><td>All other ffn_down</td><td>IQ4_KSS</td><td>Balanced 4-bit for the bulk of the feed-forward network.</td></tr>
	<tr><td>blk.50–59 ffn_gate / ffn_up</td><td>IQ4_KSS</td><td>Late FFN layers kept at 4-bit to protect output style.</td></tr>
	<tr><td>blk.0–19 ffn_gate / ffn_up</td><td>IQ4_KSS</td><td>Early FFN layers similarly balanced.</td></tr>
	<tr><td>All remaining tensors</td><td>IQ4_KS</td><td>Default fallback for maximum compression at 4-bit.</td></tr>
	</tbody>
	</table>
	</div>

	<div class="mx-h2">The V3 Calibration Dataset</div>

	<div class="mx-text">
	The importance matrix for this quantization was not generated lightly. It required over four hours of dedicated compute time, processing 300 chunks of domain-specific data at a 1024-token context window across 150 cycles.
	</div>

	<div class="mx-text">
	Crucially, this imatrix was built using a custom <strong>V3 Calibration Dataset</strong>. We took the <a href="https://huggingface.co/datasets/OpenIntelligenceNet/Uncensored-Imatrix-Calibration-Mixed-Data">OpenIntelligenceNet Mixed Data</a> and completely stripped out all Polyglot, Chain-of-Thought, and Factual knowledge data. What remains is a laser-focused dataset containing <strong>only Uncensored and Roleplay data</strong>. This ensures the quantization weights are perfectly optimized for creative and unfiltered inference, rather than wasting precision on general factual recall.
	</div>

	<div class="mx-dataset-bar">
	<div class="mx-dataset-segment mx-seg-1">40%</div>
	<div class="mx-dataset-segment mx-seg-2">25%</div>
	<div class="mx-dataset-segment mx-seg-3">20%</div>
	<div class="mx-dataset-segment mx-seg-4">15%</div>
	</div>

	<div class="mx-dataset-legend">
	<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-1"></div> 40% v2_stripped</div>
	<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-2"></div> 25% DPO Roleplay (ath)</div>
	<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-3"></div> 20% Character Cards (gry)</div>
	<div class="mx-legend-item"><div class="mx-legend-dot mx-seg-4"></div> 15% SynthRP (epi)</div>
	</div>

	<div class="mx-text" style="font-size: 0.9rem; color: #64748b;">
	The v2_stripped source consists of the Uncensored Alpaca Blend and Multi-Turn Roleplay & Fiction. The ath source provides DPO Pairs for NSFW roleplay. The gry source is Sonnet 3.5 character card roleplay. The epi source is synthetic roleplay generations.
	</div>

	<div class="mx-callout">
	<p><strong>Competitive Positioning:</strong> MaxOptimaV2 targets the same niche as <span class="mx-code">IQ4_KS</span> but with a finer-grained weight distribution. In practice, it offers comparable or slightly better roleplay coherence than <span class="mx-code">Q4_K_M</span> while producing a smaller file. It is not a universal upgrade. It is a purpose-built tool for creative and uncensored inference.</p>
	</div>

	<div class="mx-h2">Compatibility</div>
	<div class="mx-text">
	<strong>MaxOptimaV2 is exclusive to ik_llama.cpp.</strong> It will not load in mainline llama.cpp. If you require mainline compatibility or standard formats like BF16, please refer to my previous repository: <a href="https://huggingface.co/Austriani/Split-Untied-31B-ik_llama.cpp">Austriani/Split-Untied-31B-ik_llama.cpp</a>.
	</div>

	<div class="mx-footer">
	<div>Austriani // Split-Untied-31B MaxOptimaV2</div>
	<div>Base: Blazed-Forge/Split-Untied-31B // V3 Calibration</div>
	</div>

	</div>
	</div>