Qwen-AgentWorld-35B-A3B-MTP-Uncensored (无思考 · MTP)(No Think · MTP)

Qwen-AgentWorld-35B-A3B-MTP-Uncensored-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📅 2026-07-08
在 HuggingFace 查看模型 →View on HuggingFace →
能力上限Max Score
75.0
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
75.0
−0 重试扣分−11 retry penalty
ToolCall-15
83
12/15 通过 · 80%12/15 passed · 80% · 重试Retry -0
BugFind-15
83
12/15 通过 · 80%12/15 passed · 80% · 重试Retry -0
HermesAgent-20
63
9/20 通过 · 45%9/20 passed · 45% · 重试Retry -0
📊 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

  • ToolCall 83 分，11/15 题一遍过（0 重试扣分），BF-06 识别准确（88分），无思考下工具调用能力保持稳定ToolCall scored 83, with 11/15 questions first-try (0 retry penalty); BF-06 correctly identified (88 pts); tool calling remains stable without thinking mode.
  • BugFind 83 分，12/15 题一遍过，Trap 题（BF-03/10）和其他简单题识别与思考版接近BugFind scored 83, 12/15 questions first-try; Trap questions (BF-03/10) and other simple questions are close to the thinking version.
  • 重试扣分 0 —— 所有 pass 题都一遍过，要么直接答对要么直接答错，模型行为非常"诚实"Retry penalty = 0 — All passed questions went through in one shot; the model is either right or wrong, behavior is very "honest".
  • 速度优势：相比思考版省去推理步骤，预计输出速度更快（待实测）Speed advantage: Skips reasoning steps vs thinking version, expected to be faster (TBD).

⚠️ 不足Weaknesses

  • HermesAgent 仅 63 分（9/20 通过 + 3 部分分），相比思考版（79）下降 16 分，Agent 复杂场景能力严重退化HermesAgent only 63 points (9/20 passed + 3 partial), a 16-point drop from thinking version (79); complex Agent scenarios severely degraded.
  • HA-02 记忆管理（20分）、HA-04 记忆召回（0分）、HA-06 进程管理（0分）—— 记忆和执行类任务集体失分HA-02 memory mgmt (20), HA-04 memory recall (0), HA-06 process mgmt (0) — memory and execution tasks collectively failed.
  • TC-12（安全拒绝）、TC-14（错误恢复）、TC-06（多语言）失分；BF-15（环境）和 Trap 题（BF-03/10）依然失败Lost marks on TC-12 (safety rejection), TC-14 (error recovery), TC-06 (multilingual); BF-15 (environment) and Trap questions (BF-03/10) still fail.

📋 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
  • 采样参数Sampling Parameters — temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.5（无思考No Think）
  • 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

Qwen-AgentWorld-35B MTP Uncensored 无思考版 能力上限 75.0 分，实用得分 75.0 分（0 重试扣分）。35B MoE（3B 激活），16.6 GB，62.6 t/s。ToolCall 12/15 一遍过，BugFind 12/15 一遍过，HermesAgent 9/20 通过 + 3 部分分。Qwen-AgentWorld-35B MTP Uncensored no-thinking edition: Max Score 75.0, Effective Score 75.0 (0 retry penalty). 35B MoE (3B activated), 16.6 GB, 62.6 t/s. ToolCall 12/15 first-try, BugFind 12/15 first-try, HermesAgent 9/20 passed + 3 partial.

一句话评价 TL;DR：无思考版比思考版（85.6/81.6）退步 10.6 分，HermesAgent 退步最严重（-16）。TL;DR: No-thinking version is 10.6 points behind thinking version (85.6/81.6), with HermesAgent taking the biggest hit (-16).