What is cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF?

--- language: - en - zh base_model: - Qwen/Qwen3.5-9B - agentscope-ai/QwenPaw-Flash-9B tags: - heretic - abliteration - uncensored - mtp - speculative-decoding - qwen3.5 - gguf - benchlocal - benchmark - agent - tool-call license: apache-2.0 --- <div style="position: absolute; top: 50%; left: -15px; width: 60px; height: 60px; background: #ffedd5; border-ra…

What license applies to cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF?

License: apache-2.0. Verify terms on Hugging Face before commercial use.

How do I run cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF locally?

Download a GGUF file from this page and load it in guIDE or llama.cpp. Pipeline task: text-generation.

Model Intelligence Sheet

cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF overview

Q: What license applies to cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF?

License: apache-2.0. Verify terms on Hugging Face before commercial use.

Q: How do I run cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF locally?

Download a GGUF file from this page and load it in guIDE or llama.cpp. Pipeline task: text-generation.

Q: How much VRAM or disk space does cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF need?

Runs locally from ~879.0 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).

language: en zh base model: Qwen/Qwen3.5 9B agentscope ai/QwenPaw Flash 9B tags: heretic abliteration uncensored mtp speculative decoding qwen3.5 gguf benchloc…

ggufhereticabliterationuncensoredmtpspeculative-decodingqwen3.5benchlocalbenchmarkagenttool-callenzhbase_model:Qwen/Qwen3.5-9Bbase_model:quantized:Qwen/Qwen3.5-9Blicense:apache-2.0region:us

Runs locally from ~879.0 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).

Downloads

Likes

Pipeline

—

Author

cvgro

Repository Files & Downloads

5 GGUF files detected

Direct downloads for local inference

File	Type	Quantization	Size	Link
QwenPaw-Flash-9B-heretic-MTP-F16.gguf	GGUF	F16	17.14 GB	Download
QwenPaw-Flash-9B-heretic-MTP-Q4_K_M.gguf	GGUF	Q4_K_M	5.38 GB	Download
QwenPaw-Flash-9B-heretic-MTP-Q6_K.gguf	GGUF	Q6_K	7.04 GB	Download
QwenPaw-Flash-9B-heretic-MTP-Q8_0.gguf	GGUF	Q8_0	9.11 GB	Download
mmproj-QwenPaw-Flash-9B-heretic-BF16.gguf	GGUF	BF16	879.0 MB	Download

Model Details

Model ID	cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF
Author	cvgro
Pipeline	—
License	apache-2.0
Base model	Qwen/Qwen3.5-9B,agentscope-ai/QwenPaw-Flash-9B
Last modified	2026-06-29T00:50:47.000Z

Model README

---

language:

base_model:

Qwen/Qwen3.5-9B
agentscope-ai/QwenPaw-Flash-9B

tags:

heretic
abliteration
uncensored
mtp
speculative-decoding
qwen3.5
gguf
benchlocal
benchmark
agent
tool-call

license: apache-2.0

---

</div>

<h1 style="margin: 0 0 8px 0; font-size: 32px; font-weight: 700; color: #1d1d1f; letter-spacing: -0.5px; border: none; position: relative; z-index: 1;">QwenPaw-Flash-9B-heretic-MTP</h1>

<p style="margin: 8px 0 0 0; font-size: 14px; position: relative; z-index: 1;"><span style="color: #86868b;">English</span> | <a href="https://huggingface.co/SC117/QwenPaw-Flash-9B-heretic-MTP-GGUF/blob/main/README_zh.md" style="color: #007aff; text-decoration: none;">📖 中文文档</a></p>

</div>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">QwenPaw-Flash-9B-heretic non-MTP Version: <a href="https://huggingface.co/SC117/QwenPaw-Flash-9B-heretic-GGUF" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">QwenPaw-Flash-9B-heretic-GGUF</a></p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>🏆 BenchLocal Total: 4035/5000 (80.7%) — MTP Speculative Decoding Injected</b><br></p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><i>Uncensored · Abliterated · Agent-Optimized · 1.7-4.1× Speedup</i></p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Uncensored version of <b>QwenPaw-Flash-9B</b>, processed with <a href="https://github.com/p-e-w/heretic" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">Heretic</a> v1.3.0 abliteration, with <b>MTP (Multi-Token Prediction) head</b> weights injected from the original Qwen3.5-9B base model.</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">By reconstructing the MTP speculative decoding head — which was stripped during the QwenPaw fine-tuning process — this model achieves <b>up to 4.1× inference speedup</b> on real agent benchmarks while maintaining or improving accuracy.</p>

</div>

<div style="background: linear-gradient(135deg, #ff6b35 0%, #f7931e 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📊</span> 🏆 BenchLocal Benchmarks (With MTP)</div>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #64748b; font-style: italic; border-left: 3px solid #ffedd5; padding-left: 12px;"><b>Test Environment</b>: NVIDIA RTX 5070 Ti (16GB) · llama.cpp (turboquant build, <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type draft-mtp</code>) · Q6_K quant</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #64748b; font-style: italic; border-left: 3px solid #ffedd5; padding-left: 12px;"><b>Framework</b>: <a href="https://github.com/stevibe/benchlocal" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">BenchLocal</a> — local model agent evaluation suite</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #64748b; font-style: italic; border-left: 3px solid #ffedd5; padding-left: 12px;"><b>Methodology</b>: Each scenario run <b>once</b>, no retries, no second attempts</p>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Benchmark</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Score</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Accuracy</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Results</th>

</tr></thead><tbody>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>ToolCall-15</b> 🛠️</td>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>1.4×</b> faster</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>HermesAgent-20</b> 🤖</td>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>1.17×</b> faster</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>BugFind-15</b> 🐛</td>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>4.1×</b> faster</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>Total</b></td>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>1.9×</b> faster</td>

</tr>

</tbody></table>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">Comparison: With vs Without MTP</h3>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Benchmark</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Without MTP</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Δ Score</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Δ Speed</th>

</tr></thead><tbody>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>ToolCall-15</b> 🛠️</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>HermesAgent-20</b> 🤖</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>BugFind-15</b> 🐛</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>Total</b></td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>Total Time</b></td>

</tr>

</tbody></table>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🛠️ ToolCall-15 — Tool Calling Stability (100%, +6.7 pts)</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">MTP speculative decoding eliminated the single failure (TC-05: Relative date/time parsing, which previously scored 0). All 15 scenarios now pass perfectly.</p>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Result</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Scenario</th>

</tr></thead><tbody>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">Simple / Multi / Nested / Type conversion</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">Relative date/time parsing ← <b>fixed by MTP</b></td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">All remaining scenarios</td>

</tr>

</tbody></table>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🤖 HermesAgent-20 — Complex Agent Tasks (75.3%, −1.9 pts)</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">MTP decoding introduces minor noise in long-chain reasoning scenarios (~40pt drop), likely because draft tokens occasionally derail the generation path in multi-step planning tasks. However, the speed gain (1.17×) and the fact that the drop is within noise range (single-run variance was 255pts for Qwopus MTP) makes this an acceptable trade-off.</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🐛 BugFind-15 — Code Debugging (68.7%, +6.8 pts)</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Significant improvement — MTP's faster decoding effectively prevents timeout failures (BF-12 previously hit 300s timeout, now completes in time) and the draft context helps maintain debugging focus.</p>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Without MTP</th>

</tr></thead><tbody>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">❌ 0 (timeout)</td>

</tr>

<tr>

</tr>

<tr>

</tr>

<tr>

</tr>

</tbody></table>

</div>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Multi-Token Prediction (MTP) is a speculative decoding technique where a small "draft head" predicts multiple future tokens in parallel. The main model then verifies these drafts in a single forward pass, accepting correct predictions for up to 2-4× speedup in practice.</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">Injection Method</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">The original Qwen3.5-9B base model ships with a 4-layer MTP head (~243M params) in its architecture configuration. During QwenPaw fine-tuning, the MTP head weights were stripped (only the config placeholder <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">mtp_num_hidden_layers: 1</code> remained, but no actual tensors existed in the safetensors).</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>Recovery process:</b></p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>Total injected parameters:</b> 243.3M (2.7% of main model)</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>MTP acceptance rate (draft-n-max=2):</b> ~50% (1083 accepted / 2166 generated across all benchmarks)</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">Why This Works</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">The MTP head is a lightweight 4-layer MLP decoder that maps the main model's last hidden state to future token logits. It sits entirely in speculative decoding space — the main model's weights are unchanged, so no fine-tuning or retraining is needed. The head simply needs to exist with compatible dimensions for llama.cpp's <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type draft-mtp</code> to activate.</p>

</div>

<div style="background: linear-gradient(135deg, #ff6b35 0%, #f7931e 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>⚡</span> Comparison: QwenPaw MTP vs Other Models</div>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Model</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Total</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">ToolCall-15</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">HermesAgent-20</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">BugFind-15</th>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Total Time</th>

</tr></thead><tbody>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>🐾 QwenPaw MTP 9B</b></td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🐾 QwenPaw 9B (no MTP)</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🧠 Qwopus 9B MTP</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🧠 Qwen 35B Thinking ON</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">⚡ Qwen 35B Thinking OFF</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🔮 Gemma 4 26B</td>

</tr>

</tbody></table>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>QwenPaw MTP wins</b> on 2/3 benchmarks + total score + total time. The only benchmark it loses is BugFind-15 (to Qwopus MTP), but Qwopus suffers from severe instability (255pt variance on HermesAgent-20, with a worst-case 6.2min timeout).</p>

</div>

<ul style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Base model: <a href="https://huggingface.co/agentscope-ai/QwenPaw-Flash-9B" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">QwenPaw-Flash-9B</a> (Qwen3.5-9B fine-tuned for autonomous agent scenarios)</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP head source: <a href="https://huggingface.co/Qwen/Qwen3.5-9B" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">Qwen/Qwen3.5-9B</a> (original base model, layer 32 MTP head)</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Tool: Heretic v1.3.0 (automatic directional ablation)</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Best trial: #194 / 230 trials (abliteration)</li></ul>

</div>

<p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">direction_index = 21.13

attn.o_proj.max_weight = 1.42

attn.o_proj.max_weight_position = 21.72

attn.o_proj.min_weight = 1.11

attn.o_proj.min_weight_distance = 18.14

mlp.down_proj.max_weight = 1.48

mlp.down_proj.max_weight_position = 21.23

mlp.down_proj.min_weight = 1.47

mlp.down_proj.min_weight_distance = 17.47</p>

</div>

<ul style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Type: Qwen3_5ForConditionalGeneration (multimodal with vision encoder) + MTP spec head</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Main model parameters: ~9B</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP head parameters: ~243M (2.7% overhead)</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Layers: 32 (hybrid: Gated DeltaNet + Gated Attention) + 4 MTP decoder layers</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Context length: 262,144 tokens</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Speculative decoding: <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type draft-mtp</code> with <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-draft-n-max 2</code></li></ul>

</div>

<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Notes</th>

</tr></thead><tbody>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">QwenPaw-Flash-9B-heretic-MTP-Q8_0.gguf</code></td>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">High quality, near lossless</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">✅ <b>Recommended, best value</b></td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">Compact</td>

</tr>

<tr>

<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">Vision encoder (multimodal) — same as non-MTP version</td>

</tr>

</tbody></table>

</div>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">--spec-type draft-mtp</p>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">--spec-draft-n-max 2</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">llama.cpp (with MTP speculative decoding)</h3>

llama-server -m QwenPaw-Flash-9B-heretic-MTP-Q6_K.gguf \

-ngl 99 -fa on -c 8192 \

--spec-type draft-mtp --spec-draft-n-max 2 \

--host 0.0.0.0 --port 8088

Or with CLI

llama-cli -m QwenPaw-Flash-9B-heretic-MTP-Q6_K.gguf \

-ngl 99 -fa on -c 8192 \

--spec-type draft-mtp --spec-draft-n-max 2 \

-p "Write a Python script to..."</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">llama.cpp (without MTP, fallback)</h3>

llama-server -m QwenPaw-Flash-9B-heretic-MTP-Q6_K.gguf \

-ngl 99 -fa on -c 8192 \

--host 0.0.0.0 --port 8088</p>

<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">LM Studio</h3>

<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Load the GGUF file directly. For MTP speculative decoding, LM Studio must support <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type</code> — if not, the model functions as a standard 9B model.</p>

</div>

<ol style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Safety filters have been significantly reduced via abliteration</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">KL divergence is only 0.0225 — minimal impact on model intelligence</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">The original model supports multimodal (vision); GGUF versions require the mmproj file from the non-MTP release</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">BenchLocal scores measured at <b>Q6_K</b> on RTX 5070 Ti 16GB with llama.cpp (turboquant). Each scenario was run <b>once with no retries</b></li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP acceptance rate of ~50% at draft-n-max=2 means ~25-40% wall-clock speedup on short prompts, and up to 4× on long-generation tasks (debugging, code writing)</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">BugFind-15 saw the largest improvement (4.1×) because debugging tasks are generation-heavy — more tokens, more drafts accepted</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">The MTP head is a <b>lossless copy</b> from the original Qwen3.5-9B — no training was involved, simply weight injection</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Agent-heavy scenarios (HermesAgent-20) see the least MTP benefit because short-turn interactions don't give the draft head enough runway</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Please use responsibly</li></ol>

</div>

<ul style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;"><a href="https://github.com/p-e-w/heretic" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">Heretic</a> — Automated censorship removal</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;"><a href="https://huggingface.co/agentscope-ai/QwenPaw-Flash-9B" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">agentscope-ai/QwenPaw-Flash-9B</a> — Base model</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;"><a href="https://huggingface.co/Qwen/Qwen3.5-9B" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">Qwen/Qwen3.5-9B</a> — MTP head source</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;"><a href="https://github.com/ggml-org/llama.cpp" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">llama.cpp</a> — GGUF quantization and inference</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;"><a href="https://github.com/stevibe/benchlocal" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">BenchLocal</a> — Local model agent evaluation suite</li></ul>

</div>

Run cvgro/QwenPaw-Flash-9B-heretic-MTP-GGUF with guIDE

Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.

Download guIDE → · Browse 524k+ models · Compare models

Source: Hugging Face · Compare models