Qwen3.6-35B-A3B-uncensored-MTP (思考 · MTP)(Thinking · 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
89.5
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
81.5
−8 重试扣分−8 retry penalty
ToolCall-15
97
14/15 通过 · 93%14/15 passed · 93% · 重试Retry -2
BugFind-15
88
13/15 通过 · 87%13/15 passed · 87% · 重试Retry -0
HermesAgent-20
85
14/20 通过 · 70%14/20 passed · 70% · 重试Retry -6
📊 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 (9)❌ Failed (9)
#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

  • ToolCall 97 分，仅 TC-11 部分失败（50分），其余 14 题通过，工具调用基础扎实ToolCall scored 97, with only TC-11 failing (50 points); passed the remaining 14 questions, demonstrating solid tool-calling fundamentals.
  • BugFind 88 分（12/15 通过），Trap 题（BF-03、BF-10）成功识别，展现了较强的代码判断力BugFind 88 points (12/15 passed); Trap questions (BF-03, BF-10) successfully identified, demonstrating strong code judgment ability.
  • HermesAgent 85 分（14/20 通过），HA-10 首次通过（40分），HA-19 后提升至 80 分，多轮推理能力有实质性收益HermesAgent scored 85 (14/20 passed). HA-10 passed for the first time (40 points), and improved to 80 later; multi-turn reasoning showed substantial gains.
  • 仅 8 次（TC-1、HA-2、HA-2、HA-1、HA-3），整体较为稳定Only 8 instances (TC-1, HA-2, HA-2, HA-1, HA-3); overall performance was quite stable.

⚠️ 不足Weaknesses

  • HermesAgent 仍有 6 题失败（HA-04/07/10/14/16/19），Agent 场景（记忆召回、批量处理、消息投递）仍存在短板HermesAgent still failed 6 questions (HA-04/07/10/14/16/19); Agent scenarios (memory recall, batch processing, message delivery) remain weak points.
  • TC-11 部分失败（50分），不必要的工具调用问题未能解决TC-11 partially failed (50 points); unnecessary tool calling issue unresolved.
  • BF-10 虽然通过但仅得 40 分（前），说明 Trap 场景下仍需多次尝试才能定位BF-10 passed but scored only 40 points (early), indicating that multiple attempts are still needed to locate the trap in Trap scenarios.

📋 测试环境Test Environment

  • — RTX 5070 Ti 16GB + 128GB RAM，MoE 部分专家层 offload 到 CPU— RTX 5070 Ti 16GB + 128GB RAM, MoE expert layers offloaded to CPU.
  • 推理后端Inference Backend — llama.cpp
  • 采样参数 — temperature=1.0, top_p=0.95, top_k=20, repetition_penalty=1.5Sampling parameters — temperature=1.0, top_p=0.95, top_k=20, repetition_penalty=1.5
  • 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 思考版 89.5 分， 81.5 分（ 8 次）。35B MoE（3B 激活），16.6 GB，62.6 t/s。ToolCall 稳健（97），BugFind 通过率 80%，HermesAgent 在 Agent 场景下中规中矩。思考模式在 Trap 场景和多轮推理上有明显提升。Qwen3.6-35B MTP Thinking Edition: 89.5 points, 81.5 points (8 runs). 35B MoE (3B activated), 16.6 GB, 62.6 t/s. Robust ToolCall (97), BugFind pass rate 80%, HermesAgent performs average in Agent scenarios. Thinking mode shows significant improvement in Trap scenarios and multi-turn reasoning.

一句话评价 TL;DR：思考模式让 Trap 题和 Agent 复杂场景有了明显改善——仅 8 次， 81.5。TL;DR: The thinking mode brought clear improvements to Trap questions and complex Agent scenarios—only 8 instances, scoring 81.5.