docs: complete Chinese translation

README_zh.md

@@ -28,35 +28,35 @@ license: apache-2.0
     <span style="background: #007aff; color: white; font-size: 11px; font-weight: 600; padding: 5px 14px; border-radius: 20px;">MTP</span><span style="background: #af52de; color: white; font-size: 11px; font-weight: 600; padding: 5px 14px; border-radius: 20px;">GGUF</span>
   </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;"><a href="https://huggingface.co/SC117/QwenPaw-Flash-9B-heretic-MTP-GGUF/blob/main/README.md" style="color: #007aff; text-decoration: none;">📖 English</a> | <span style="color: #86868b;">中文</span></p>
+<p style="margin: 8px 0 0 0; font-size: 14px; position: relative; z-index: 1;"><a href="https://huggingface.co/SC117/QwenPaw-Flash-9B-heretic-MTP-GGUF/blob/main/README.md" style="color: #007aff; text-decoration: none;">📖 English</a> | <span style="color: #86868b;">中文</span> style="color: #007aff; text-decoration: none;">📖 中文文档</a></p>
 </div>
 </div>
 
 <div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; display: flex; flex-direction: column; gap: 20px; margin-bottom: 30px;">
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
       <div style="padding: 16px;">
-<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;">QwenPaw-Flash-9B-heretic 非 MTP 版本： <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;"><p align="center"></p>
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>🏆 BenchLocal 总计: 4035/5000 (80.7%) — MTP 投机解码 Injected</b><br></p>
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><i>无审查 · 已消融 · Agent 优化 · 1.7-4.1× 加速比</i></p>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>🏆 BenchLocal 总分: 4035/5000 (80.7%) — MTP 投机解码已注入</b><br></p>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><i>无审查 · 已消融 · Agent 优化 · 1.7-4.1× 加速</i></p>
 <p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"></p></p>
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">以下模型的无审查版本： <b>QwenPaw-Flash-9B</b>, 使用 <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（多 token 预测）) 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>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>QwenPaw-Flash-9B</b> 的无审查版本，使用 <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 消融处理，并从原始 Qwen3.5-9B 基座模型注入了 <b>MTP（Multi-Token Prediction）头</b>权重。</p>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">通过重建在 QwenPaw 微调过程中被剥离的 MTP 投机解码头，本模型在真实 Agent 基准测试中实现了<b>最高 4.1× 推理加速</b>，同时保持或提升了准确率。</p>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
-      <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 基准测试 (With MTP)</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 基准测试（��用 MTP）</div>
       <div style="padding: 16px;">
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #64748b; font-style: italic; border-left: 3px solid #ffedd5; padding-left: 12px;"><b>测试环境</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>测试框架</b>: <a href="https://github.com/stevibe/benchlocal" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">BenchLocal</a> — 本地模型 Agent 评估套件</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>测试方法</b>: 每个场景运行 <b>一次</b>, 无重试，无二次尝试</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>测试环境</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>测试框架</b>： <a href="https://github.com/stevibe/benchlocal" target="_blank" style="color: #c2410c; text-decoration: none; font-weight: 700;">BenchLocal</a> — 本地模型 Agent 评估套件</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>测试方法</b>：每个场景运行<b>一次</b>，无重试，无二次尝试</p>
 <table style="width: 100%; border-collapse: collapse; font-size: 13px;"><thead><tr style="background: rgba(255,107,53,0.05);">
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">基准测试</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">得分</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">准确率</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">结果</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">耗时</th>
-<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">vs No-MTP</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">对比无 MTP</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>
@@ -83,7 +83,7 @@ license: apache-2.0
 <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>总计</b></td>
+<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>4035/5000</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>80.7%</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">36✅ 3⚠️ 11❌</td>
@@ -91,11 +91,11 @@ license: apache-2.0
 <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>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">对比：有 MTP vs 无 MTP</h3>
 <table style="width: 100%; border-collapse: collapse; font-size: 13px;"><thead><tr style="background: rgba(255,107,53,0.05);">
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">基准测试</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;">With MTP</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">无 MTP</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">有 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>
@@ -121,51 +121,51 @@ license: apache-2.0
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>4.1×</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>总计</b></td>
+<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>3873/5000 (77.5%)</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>4035/5000 (80.7%)</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>+162 pts</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></td>
 </tr>
 <tr>
-<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>总计 Time</b></td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>总耗时</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>14.7 min</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>7.8 min</b></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">—</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></td>
 </tr>
 </tbody></table>
-<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🛠️ ToolCall-15 — 工具调用稳定性 (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: 相对日期/时间解析, which previously scored 0). All 15 scenarios now pass perfectly.</p>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🛠️ ToolCall-15 — 工具调用稳定性 (100%, +6.7 分)</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">MTP 投机解码消除了唯一的失败项（TC-05：相对日期/时间解析，之前得分为 0）。全部 15 个场景现在完美通过。</p>
 <table style="width: 100%; border-collapse: collapse; font-size: 13px;"><thead><tr style="background: rgba(255,107,53,0.05);">
-<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">TC-ID</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">编号</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">结果</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">场景</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;">TC-01–TC-04</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">✅</td>
-<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>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">简单 / 多参数 / 嵌套 / 类型转换</td>
 </tr>
 <tr>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">TC-05</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">✅</td>
-<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">相对日期/时间解析 ← <b>fixed by MTP</b></td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">相对日期/时间解析 ← <b>已被 MTP 修复</b></td>
 </tr>
 <tr>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">TC-06–TC-15</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">✅</td>
-<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">All remaining scenarios</td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">所有剩余场景</td>
 </tr>
 </tbody></table>
-<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🤖 HermesAgent-20 — 复杂 Agent 任务 (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>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🤖 HermesAgent-20 — 复杂 Agent 任务 (75.3%, −1.9 分)</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">MTP 解码在长链推理场景中引入了轻微噪声（约 40 分下降），可能是因为 draft token 偶尔在多步规划任务中偏离生成路径。不过，速度提升（1.17×）以及下降幅度在噪声范围内（Qwopus MTP 单次运行方差为 255 分）使得这是一个可接受的权衡。</p>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">🐛 BugFind-15 — 代码调试 (68.7%, +6.8 分)</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">显著提升 — MTP 更快的解码有效防止了超时失败（BF-12 之前触发 300 秒超时，现在按时完成），draft 上下文有助于保持调试焦点。</p>
 <table style="width: 100%; border-collapse: collapse; font-size: 13px;"><thead><tr style="background: rgba(255,107,53,0.05);">
-<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">BF-ID</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;">With MTP</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">编号</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">无 MTP</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">有 MTP</th>
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">Δ</th>
 </tr></thead><tbody>
 <tr>
@@ -262,23 +262,23 @@ license: apache-2.0
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
-      <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> MTP 投机解码</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> MTP Speculative Decoding</div>
       <div style="padding: 16px;">
-<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">What is MTP?</h3>
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Multi-Token Prediction（多 token 预测） (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>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">什么是 MTP？</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">Multi-Token Prediction（MTP）是一种投机解码技术，通过一个小型「draft 头」并行预测多个未来 token。主模型随后在单次前向传播中验证这些预测，接受正确的预测以实现 2-4× 的实际加速。</p>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">注入方法</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">原始 Qwen3.5-9B 基座模型在架构配置中附带了一个 4 层 MTP 头（约 243M 参数）。在 QwenPaw 微调过程中，MTP 头权重被剥离（仅保留了配置占位符 <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> ，但 safetensors 中没有实际张量）。</p>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>恢复过程：</b></p>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
 <ol style="margin: 0 0 12px 0; padding-left: 20px;"></ol>
-<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>总计 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>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>注入参数总量：</b>243.3M（主模型的 2.7%）</p>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>MTP 接受率 (draft-n-max=2)：</b>约 50%（所有基准测试中 1083 次接受 / 2166 次生成）</p>
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">为什么有效</h3>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">MTP 头是一个轻量级 4 层 MLP 解码器，将主模型的最后隐藏状态映射到未来 token 的 logits。它完全位于投机解码空间中 — 主模型权重不变，因此无需微调或重新训练。MTP 头只需以兼容的维度存在，llama.cpp 的 <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type draft-mtp</code> 即可激活。</p>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
@@ -286,11 +286,11 @@ license: apache-2.0
       <div style="padding: 16px;">
 <table style="width: 100%; border-collapse: collapse; font-size: 13px;"><thead><tr style="background: rgba(255,107,53,0.05);">
 <th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">模型</th>
-<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">总计</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;">总计 Time</th>
+<th style="padding: 8px 10px; border-bottom: 2px solid #ff6b35; text-align: left; color: #c2410c; font-weight: bold;">总耗时</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>
@@ -301,7 +301,7 @@ license: apache-2.0
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>7.8min</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>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🐾 QwenPaw 9B（无 MTP）</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">3873</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">93.3%</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;"><b>77.2%</b> 🥇</td>
@@ -317,7 +317,7 @@ license: apache-2.0
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">21.3min ⚠️</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 思考模式 开</td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">🧠 Qwen 35B 思考模式开</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">1445 (HA only)</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">—</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">72.3%</td>
@@ -325,7 +325,7 @@ license: apache-2.0
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">7.0min</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 思考模式 关</td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">⚡ Qwen 35B 思考模式关</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">1370 (HA only)</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">—</td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">68.5%</td>
@@ -341,13 +341,13 @@ license: apache-2.0
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">18.6min</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>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;"><b>QwenPaw MTP 胜出</b>2/3 基准测试 + 总分 + 总耗时。唯一输掉的基准是 BugFind-15（输给 Qwopus MTP），但 Qwopus 存在严重不稳定性（HermesAgent-20 方差 255 分，最差情况 6.2 分钟超时）。</p>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
       <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> Model Description</div>
       <div style="padding: 16px;">
-<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>
+<ul style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">基座模型**： <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 针对自主 Agent 场景微调）</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP 头来源**： <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> （原始基座模型，第 32 层 MTP 头）</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">工具**：Heretic v1.3.0（自动定向消融）</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">最佳试验**：#194 / 230 次试验（消融）</li></ul>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
@@ -367,7 +367,7 @@ mlp.down_proj.min_weight_distance = 17.47</p>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
       <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> Architecture</div>
       <div style="padding: 16px;">
-<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>
+<ul style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">类型**：Qwen3_5ForConditionalGeneration（多模态，含视觉编码器）+ MTP 投机解码头</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">主模型参数量**：~9B</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP 头参数量**：~243M（2.7% 额外开销）</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">层数**：32（混合：Gated DeltaNet + Gated Attention）+ 4 层 MTP 解码器</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">上下文长度**：262,144 tokens</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">投机解码**： <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type draft-mtp</code> 配合 <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>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
@@ -381,22 +381,22 @@ mlp.down_proj.min_weight_distance = 17.47</p>
 <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;">~9.2GB</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>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">高质量，近乎无损</td>
 </tr>
 <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-Q6_K.gguf</code></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">~7.1GB</td>
-<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>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">✅ <b>推荐，最佳性价比</b></td>
 </tr>
 <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-Q4_K_M.gguf</code></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">~5.4GB</td>
-<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">Compact</td>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">紧凑</td>
 </tr>
 <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;">mmproj-BF16</code></td>
 <td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">~880MB</td>
-<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>
+<td style="padding: 7px 10px; border-bottom: 1px solid rgba(128,128,128,0.15); color: #334155; text-align: left;">视觉编码器（多模态）— 与非 MTP 版本相同</td>
 </tr>
 </tbody></table>
       </div>
@@ -406,37 +406,37 @@ mlp.down_proj.min_weight_distance = 17.47</p>
       <div style="padding: 16px;">
 <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>
-<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;"># Start server with MTP enabled
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">llama.cpp (配合 MTP speculative decoding)</h3>
+<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;"># Start server 配合 MTP enabled
 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
+# Or 配合 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>
-<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;"># The model works as a normal GGUF too — just omit spec args
+<h3 style="margin: 16px 0 8px 0; font-size: 14px; color: #1e293b; font-weight: 700;">llama.cpp (配合out MTP, fallback)</h3>
+<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;"># 模型也可作为普通 GGUF 使用 — 只需省略投机解码参数
 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>
+<p style="margin: 0 0 12px 0; font-size: 13px; color: #334155; line-height: 1.7;">直接加载 GGUF 文件。如需 MTP 投机解码，LM Studio 需要支持 <code style="background: #f8fafc; padding: 2px 6px; border-radius: 4px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b;">--spec-type</code> — 如果不支持，模型将作为标准 9B 模型运行。</p>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
       <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> Notes</div>
       <div style="padding: 16px;">
-<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>一次 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>
+<ol style="margin: 0 0 12px 0; padding-left: 20px;"><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">安全过滤器已通过消融显著降低</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">KL 散度仅为 0.0225 — 对模型智能影响极小</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">原始模型支持多模态（视觉）；GGUF 版本需要非 MTP 版本的 mmproj 文件</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">BenchLocal 分数在 <b>Q6_K</b> on RTX 5070 Ti 16GB 配合 llama.cpp (turboquant). Each scenario was run <b>once 配合 no retries</b></li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP 在 draft-n-max=2 下约 50% 的接受率意味着短提示约 25-40% 的实际加速，长生成任务（调试、代码编写）最高可达 4×</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">BugFind-15 提升最大（4.1×），因为调试任务是生成密集型 — token 更多，接受的 draft 更多</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">MTP 头是从原始 Qwen3.5-9B 的<b>无损拷贝</b> — 不涉及训练，仅是权重注入</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">Agent 密集型场景（HermesAgent-20）从 MTP 获益最少，因为短轮次交互没有给 draft 头足够的发挥空间</li><li style="margin-bottom: 4px; font-size: 13px; color: #334155;">请负责任地使用</li></ol>
       </div>
     </div>
     <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);">
       <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> Acknowledgements</div>
       <div style="padding: 16px;">
-<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>
+<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> — 自动化审查移除</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> — 基座模型</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 头来源</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 量化与推理</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> — 本地模型 Agent 评估套件</li></ul>
       </div>
     </div>
 </div>