Qwen-AgentWorld-35B-A3B-APEX-I-Compact (原版 · 思考)(Original · Thinking)

Qwen-AgentWorld-35B-A3B-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
88.5
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
87.5
−1 重试扣分−8 retry penalty
ToolCall-15
100
15/15 通过 · 100%15/15 passed · 100% · 重试Retry -0
BugFind-15
87
13/15 通过 · 87%13/15 passed · 87% · 重试Retry -0
HermesAgent-20
81
13/20 通过 · 65%13/20 passed · 65% · 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 (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 100 分满分!15/15 题一遍过(0 重试),是 35B 系列里唯一的 ToolCall 满分ToolCall 100 points (perfect)! 15/15 questions first-try (0 retries), the only perfect ToolCall score in the 35B family.
  • BugFind 87 分(13/15 通过),BF-15(Go 竞态)通过,BF-06(async/await)识别准确(88分)BugFind 87 points (13/15 passed); BF-15 (Go race) and BF-06 (async/await, 88 pts) both identified correctly.
  • HermesAgent 81 分(13/20 通过 + 1 部分分),记忆管理(HA-01/02/03)、Skill 创建/更新/辅助(HA-09~12)、Cron 调度(HA-13/15)全部一次通过HermesAgent 81 points (13/20 passed + 1 partial). Memory (HA-01/02/03), Skill create/update/aux (HA-09~12), Cron (HA-13/15) all first-try.
  • 重试扣分仅 1 分(HA-08 浏览器导出),整体非常稳定,几乎所有题目一遍过Only 1 retry penalty (HA-08 browser export); overall very stable, almost all questions first-try.

⚠️ 不足Weaknesses

  • HermesAgent 仍有 6 题失败(HA-04/07/14/16/17/20)+ 1 题部分分(HA-19),Agent 复杂场景(跨平台投递、并行委派、模糊澄清)仍是能力上限HermesAgent still has 6 failures (HA-04/07/14/16/17/20) + 1 partial (HA-19); complex Agent scenarios (cross-platform delivery, parallel delegation, ambiguity clarification) remain the capability ceiling.
  • Trap 题(BF-03 Rust / BF-10 Python Red-Herring)依然完全失败(0分),35B 系列固有问题Trap questions (BF-03 Rust / BF-10 Python Red-Herring) still fail completely (0 pts); a chronic issue across the 35B family.
  • HA-04(Docker 修复召回)35分、HA-07(批处理 200 文件)30分 —— 长上下文记忆和大规模任务执行仍是短板HA-04 (Docker fix recall) 35 pts, HA-07 (batch process 200 files) 30 pts — long-context memory and large-scale task execution remain weak.

📋 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 DownloadHF: SC117

Qwen-AgentWorld-35B 原版(开思考)能力上限 88.5 分实用得分 87.5 分(仅 1 次重试)。35B MoE(3B 激活),16.6 GB,62.6 t/s。ToolCall 100 分满分,BugFind 13/15 一遍过,HermesAgent 13/20 通过 + 1 部分分。整体非常稳定,是目前 35B 系列里实用得分最高的版本。Qwen-AgentWorld-35B Original (thinking): Max Score 88.5, Effective Score 87.5 (only 1 retry). 35B MoE (3B activated), 16.6 GB, 62.6 t/s. ToolCall 100 perfect, BugFind 13/15 first-try, HermesAgent 13/20 passed + 1 partial. Overall very stable—currently the highest Effective Score in the 35B family.

一句话评价 TL;DR:原汁原味才是最强者——不加 MTP / Uncensored 反而跑出最高实用分 87.5,超过 Qwen3.6-35B 思考版 6 分!TL;DR: The original is the strongest—without MTP / Uncensored modifications, this model achieves the highest Effective Score of 87.5, beating Qwen3.6-35B thinking version by 6 points!