Qwen-AgentWorld-35B-A3B-MTP-Uncensored (思考 · MTP)(Thinking · 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📅 2026-07-07
在 HuggingFace 查看模型 →View on HuggingFace →
能力上限Max Score
85.6
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
81.6
−4 重试扣分−4 retry penalty
ToolCall-15
93
14/15 通过 · 93%14/15 passed · 93% · 重试Retry -0
BugFind-15
87
13/15 通过 · 87%13/15 passed · 87% · 重试Retry -1
HermesAgent-20
79
13/20 通过 · 65%13/20 passed · 65% · 重试Retry -3
📊 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 93 分，14/15 题一遍过（0 重试扣分），除 TC-12 外工具调用稳健可靠ToolCall scored 93, with 14/15 questions passed in one shot (0 retry penalty); tool calling is solid and reliable except TC-12.
  • BugFind 87 分（13/15 通过），仅 BF-15 重试 1 次，其余 12 题一遍过，Trap 题（BF-03、BF-10）整体识别稳定BugFind scored 87 (13/15 passed); only BF-15 retried once, the other 12 passed in one shot; Trap scenarios (BF-03, BF-10) are stable overall.
  • HermesAgent 79 分（13/20 通过 + 2 部分分），记忆管理（HA-01/02/03）、Skill 创建更新（HA-09/10/11）、Cron 调度（HA-13/14/15）一次性通过HermesAgent scored 79 (13/20 passed + 2 partial). Memory management (HA-01/02/03), Skill creation/update (HA-09/10/11), Cron scheduling (HA-13/14/15) all passed on first try.
  • 总计仅 4 次重试（BF-15、HA-03、HA-06、HA-20），整体非常稳定，几乎一次到位Only 4 retries total (BF-15, HA-03, HA-06, HA-20); overall very stable, almost all on first try.

⚠️ 不足Weaknesses

  • HermesAgent 仍有 5 题完全失败（HA-04/07/12/16/17）+ 2 题部分分（HA-08/19），Agent 复杂场景（并行委派、批处理、跨平台投递）仍是短板HermesAgent still has 5 complete failures (HA-04/07/12/16/17) + 2 partial (HA-08/19); complex Agent scenarios (parallel delegation, batch processing, cross-platform delivery) remain weak points.
  • HA-04 / HA-16 因 BenchLocal 自身运行中断失败（0分），非模型问题，但拉低了 HA 总分HA-04 / HA-16 failed (0 points) due to BenchLocal itself crashing, not model issues; however, this dragged down the HermesAgent total score.
  • TC-12（impossible request）完全失败，BugFind 的 BF-03（Rust Trap）和 BF-10（Python Red-Herring）也完全失败，说明 Trap 场景识别仍是该模型能力上限TC-12 (impossible request) failed completely; BugFind's BF-03 (Rust Trap) and BF-10 (Python Red-Herring) also failed completely, indicating Trap scenario identification remains this model's capability ceiling.

📋 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

Qwen-AgentWorld-35B MTP Uncensored 思考版 能力上限 85.6 分，实用得分 81.6 分（仅 4 次重试）。35B MoE（3B 激活），16.6 GB，62.6 t/s。ToolCall 14/15 一遍过（0 重试），BugFind 13/15 一遍过，HermesAgent 13/20 通过 + 2 部分分。整体非常稳定，是 Qwen3.6-35B 系列里最"省心"的版本之一。Qwen-AgentWorld-35B MTP Uncensored Thinking Edition: Max Score 85.6, Effective Score 81.6 (only 4 retries). 35B MoE (3B activated), 16.6 GB, 62.6 t/s. ToolCall 14/15 first-try (0 retries), BugFind 13/15 first-try, HermesAgent 13/20 passed + 2 partial. Overall very stable—one of the most "worry-free" models in the Qwen3.6-35B family.

一句话评价 TL;DR：Uncensored 变体在 Trap 题（BF-03/10）和 Agent 复杂场景（HA-17 并行委派）上仍有能力天花板，但日常任务（记忆、Skill、Cron、工具调用）几乎无重试，性价比突出。TL;DR: The Uncensored variant still hits a capability ceiling on Trap questions (BF-03/10) and complex Agent scenarios (HA-17 parallel delegation), but daily tasks (memory, Skill, Cron, tool calls) require almost no retries—excellent value for money.