Qwen3.6-35B-A3B-uncensored-MTP (无思考 · MTP)(No Think · MTP)

Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-I-Compact.gguf · Q4_K_M · 16.6 GB
🧠 MoE (3B active)📦 16.6 GB⚡ In 766 t/s · Out 62.6 t/s🎮 RTX 5070 Ti 16GB + 128GB RAM📅 2026-06-23
在 HuggingFace 查看模型 →View on HuggingFace →
能力上限Max Score
91.5
加权原始分 = TC×0.3 + BF×0.3 + HA×0.4Weighted raw score = TC×0.3 + BF×0.3 + HA×0.4
实用得分Effective Score
80.5
−11 重试扣分−11 retry penalty
ToolCall-15
97
14/15 通过 · 93%14/15 passed · 93% · 重试Retry -5
BugFind-15
96
13/15 通过 · 87%13/15 passed · 87% · 重试Retry -5
HermesAgent-20
84
15/20 通过 · 75%15/20 passed · 75% · 重试Retry -1
📊 计分规则Scoring Rules — 能力上限Max Score = 加权原始分 = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4= Weighted raw score = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4
实用得分Effective Score = 能力上限 − 重试扣分（每题首次通过不扣分，重试后通过每重试1次扣1分，重试后仍失败不参与扣分）。Max Score − retry penalty (first pass = no penalty, each retry = +1pt, failed retries excluded).

📋 全部测试结果All Test Results

全部 (50)All (50)
ToolCall (15)
BugFind (15)
HermesAgent (20)
❌ 失败 (8)❌ Failed (8)
#题目Question测试包Pack难度Difficulty结果Result得分Score重试Retrysandbox耗时Time失败类型Failure Type点评Comment

🔍 错题分析Error Analysis

🌊 📊 模型评估总结Model Evaluation Summary

✅ 优势Strengths

  • BugFind 96 分，是所有无思考中最高，代码调试能力强BugFind scored 96, the highest among all non-thinking models, with strong code debugging capabilities.
  • ToolCall 97 分，仅 TC-11 部分失败（50分），其余全通过ToolCall 97, only partial failure in TC-11 (50 points), others fully passed.
  • 与同架构无 MTP 版本相比， +1.0（90.5→91.5）， +7.0（73.5→80.5），稳定性显著提升Compared to the same architecture without MTP, it improved by +1.0 (90.5→91.5) and +7.0 (73.5→80.5), showing significantly enhanced stability.
  • 次数从 17 降到 11，MTP 让在场景下更容易成功The number of attempts decreased from 17 to 11; MTP made success more likely in this scenario.

⚠️ 不足Weaknesses

  • HermesAgent 84 分，5 题失败（HA-04/07/14/16/17），Agent 场景仍有短板HermesAgent 84, failed 5 questions (HA-04/07/14/16/17); Agent scenarios still have shortcomings.
  • TC-11 部分失败（50分），不必要的工具调用TC-11 partial failure (50 points), unnecessary tool calls.
  • BF-02（88分）和 BF-03 Trap（40分）未达满分BF-02 (88) and BF-03 Trap (40) did not achieve full marks.

📋 测试环境Test Environment

  • 硬件Hardware — RTX 5070 Ti 16GB + 128GB RAM，MoE模型Models部分专家层offload到CPUexpert layers partially offloaded to CPU
  • 推理后端Inference Backend — llama.cpp
  • 测试包Pack — ToolCall-15 / BugFind-15 / HermesAgent-20（共 50 题）— ToolCall-15 / BugFind-15 / HermesAgent-20 (total 50 questions)
  • 模型下载Model Download — HF: SC117

Qwen3.6-35B MTP 版（无思考） 91.5 分， 80.5 分（ 11 次）。35B MoE（3B 激活），16.6GB，62.6 t/s。开启 MTP 后相比无 MTP 版本，小幅提升（+1.0），大涨（+7.0），稳定性明显增强。Qwen3.6-35B MTP version (no thinking) 91.5 points, 80.5 points (11 times). 35B MoE (3B activated), 16.6GB, 62.6 t/s. After enabling MTP compared to the non-MTP version, slight improvement (+1.0), significant increase (+7.0), stability significantly enhanced.

一句话评价 TL;DR：MTP 加持后稳定性飞升——同样的 35B MoE，从 17 次降到 11 次， +7。TL;DR: With MTP support, stability skyrocketed—same 35B MoE model, reduced from 17 attempts to 11, +7.