⚠️ 非严谨学术评测，结果仅供参考与简单对比Non-strict academic evaluation — results are for reference and rough comparison only

本地模型Local Models
测试数据库Benchmark Database

基于 BenchLocal 桌面应用，在本地对真实任务进行基准测试，横向对比模型表现Based on BenchLocal desktop app, benchmarking LLMs on real-world tasks locally

当前测试环境：RTX 5070 Ti 16GB + 128GB RAM · MoE模型部分专家层offload到CPUTest environment: RTX 5070 Ti 16GB + 128GB RAM · MoE expert layers partially offloaded to CPU

排名基于能力上限（加权原始分），实用得分 = 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）Ranked by Max Score (weighted raw score), Effective Score = Max Score − retry penalty (first pass = no penalty, each retry = +1pt)

⚡ 卡片中的速度数据（如 In 725 t/s · Out 60.6 t/s）仅供参考，非真实环境！Speed data on cards (e.g. In 725 t/s · Out 60.6 t/s) is for reference only!

21
已测试模型Models Tested
50
总测试题数Total Questions
3
测试项目Test Suites

🏆 模型排行榜Model Leaderboard

能力上限Max Score
实用得分Effective Score
ToolCall
BugFind
HermesAgent
全部All
🧠 仅思考Thinking only
⚡ 仅无思考No-thinking only
1
上限Max
88.8
实用Eff.
78.8

Step-3.7-Flash-APEX-I-Mini (思考)(Thinking)

Step-3.7-Flash-APEX-I-Mini.gguf · Q3_K_M · 67.9 GB
🧠 MoE 📦 67.0 GB ⚡ In 98 t/s · Out 13.9 t/s 🎮 RTX 5070 Ti 🌡️ temp=1.0
ToolCall-15
100
BugFind-15
87
HermesAgent-20
81
🔄 上限分 88.8 → 重试 -10 → 实用 78.8🔄 Max 88.8 → retries -10 → Eff. 78.8
2026-06-21 测试tested 查看详情 →View Details →
2
上限Max
87.2
实用Eff.
60.2

Nex-N2-Mini-abliterated-APEX (思考)(Thinking)

Huihui-Nex-N2-mini-abliterated-APEX-I-Compact.gguf · Q4_K_M · 15.4 GB
🧠 MoE 📦 15.4 GB ⚡ In 321 t/s · Out 62.5 t/s 🎮 RTX 5070 Ti 🌡️ temp=0.7 top_p=0.95 top_k=40
ToolCall-15
93
BugFind-15
88
HermesAgent-20
81
🔄 上限分 87.2 → 重试 -27 → 实用 60.2🔄 Max 87.2 → retries -27 → Eff. 60.2
2026-06-20 测试tested 查看详情 →View Details → HF
3
上限Max
91.5
实用Eff.
80.5

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 📦 16.6 GB ⚡ In 766 t/s · Out 62.6 t/s 🎮 RTX 5070 Ti 🌡️ temp=0.7 top_p=0.8 top_k=20 rep=1.5
ToolCall-15
97
BugFind-15
96
HermesAgent-20
84
🔄 上限分 91.5 → 重试 -11 → 实用 80.5🔄 Max 91.5 → retries -11 → Eff. 80.5
2026-06-23 测试tested 查看详情 →View Details → HF
1
上限Max
89.5
实用Eff.
81.5

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 📦 16.6 GB ⚡ In 766 t/s · Out 62.6 t/s 🎮 RTX 5070 Ti 🌡️ temp=1.0 top_p=0.95 top_k=20 rep=1.5
ToolCall-15
97
BugFind-15
88
HermesAgent-20
85
🔄 上限分 89.5 → 重试 -8 → 实用 81.5🔄 Max 89.5 → retries -8 → Eff. 81.5
2026-06-23 测试tested 查看详情 →View Details → HF
1
上限Max
88.7
实用Eff.
82.7

Gemma-4-26B-A4B-it-qat (无思考)(No Think)

gemma-4-26B-A4B-it-qat-heretic-UD-Q4_K_XL.gguf · Q4_K_XL · 13.3 GB
🧠 MoE 4B active 📦 13.3 GB ⚡ In 832 t/s · Out 53.9 t/s 🌡️ temp=1.0 top_p=0.95 top_k=64
ToolCall-15
96.7
BugFind-15
86
HermesAgent-20
84.8
🔄 上限分 88.7 → 重试 -6 → 实用 82.7🔄 Max 88.7 → retries -6 → Eff. 82.7
2026-06-19 测试tested 查看详情 →View Details → HF
1
上限Max
91.6
实用Eff.
83.6

Gemma-4-26B-A4B-it-qat (思考)(Thinking)

gemma-4-26B-A4B-it-qat-heretic-UD-Q4_K_XL.gguf · Q4_K_XL · 13.3 GB
🧠 MoE 4B active 📦 13.3 GB ⚡ In 832 t/s · Out 53.9 t/s 🌡️ temp=1.0 top_p=0.95 top_k=64
ToolCall-15
93.3
BugFind-15
96
HermesAgent-20
87
🔄 上限分 91.6 → 重试 -8 → 实用 83.6🔄 Max 91.6 → retries -8 → Eff. 83.6
2026-06-19 测试tested 查看详情 →View Details → HF
1
上限Max
90.3
实用Eff.
85.3

Gemma-4-12B-it-heretic-QAT (思考)(Thinking)

gemma-4-12B-it-heretic-QAT-Q4_0.gguf · Q4_0 · 6.4 GB
🧠 Dense 12B 📦 6.4 GB ⚡ In 1522 t/s · Out 126.3 t/s 🌡️ QAT量化🌡️ QAT Quantization
ToolCall-15
97
BugFind-15
88
HermesAgent-20
87
🔄 上限分 90.3 → 重试 -5 → 实用 85.3🔄 Max 90.3 → retries -5 → Eff. 85.3
2026-06-21 测试tested 查看详情 →View Details → HF
-
上限Max
89.8
实用Eff.
68.8

Ornith-1.0-9B-heretic-MTP (思考)(Thinking)

Ornith-1.0-9B-heretic-MTP-Q6_K.gguf · Q6_K · 7.6 GB
🧠 Dense 9B 📦 7.6 GB ⚡ In 1420 t/s · Out 109.4 t/s 🌡️ temp=0.6 top_p=0.95 top_k=20
ToolCall-15
100
BugFind-15
94
HermesAgent-20
79
🔄 上限分 89.8 → 重试 -21 → 实用 68.8🔄 Max 89.8 → retries -21 → Eff. 68.8
2026-06-29 测试tested 查看详情 →View Details → HF
2
上限Max
88.4
实用Eff.
64.4

QwenPaw-Flash-9B-MTP (思考 · MTP)(Thinking · MTP)

QwenPaw-Flash-9B-heretic-MTP-Q6_K.gguf · Q6_K · 7.0 GB
🧠 Dense 9B 📦 7.0 GB ⚡ In 507 t/s · Out 90.3 t/s 🌡️ temp=1.0 top_p=0.95 top_k=20 rep=1.5
ToolCall-15
100
BugFind-15
84
HermesAgent-20
80
🔄 上限分 88.4 → 重试 -24 → 实用 64.4🔄 Max 88.4 → retries -24 → Eff. 64.4
2026-06-19 测试tested 查看详情 →View Details → HF
3
上限Max
87.4
实用Eff.
84.4

Qwen3.6-27B-uncensored-MTP (无思考 · MTP)(No Think · MTP)

Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved.i1-IQ3_M.gguf · IQ3_M · 11.9 GB
🧠 Dense 27B 📦 11.9 GB ⚡ In 725 t/s · Out 60.6 t/s 🌡️ temp=0.7 top_p=0.8 top_k=20 rep=1.5
ToolCall-15
96.7
BugFind-15
84.2
HermesAgent-20
82.8
🔄 上限分 87.4 → 重试 -3 → 实用 84.4🔄 Max 87.4 → retries -3 → Eff. 84.4
2026-06-19 测试tested 查看详情 →View Details →
1
上限Max
91.9
实用Eff.
73.9

Qwen3.6-27B-uncensored-MTP (思考 · MTP)(Thinking · MTP)

Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved.i1-IQ3_M.gguf · IQ3_M · 11.9 GB
🧠 Dense 27B 📦 11.9 GB ⚡ In 725 t/s · Out 60.6 t/s 🌡️ temp=1.0 top_p=0.95 top_k=20 rep=1.5
ToolCall-15
100
BugFind-15
93.3
HermesAgent-20
84.8
🔄 上限分 91.9 → 重试 -18 → 实用 73.9🔄 Max 91.9 → retries -18 → Eff. 73.9
2026-06-19 测试tested 查看详情 →View Details →
1
上限Max
94.0
实用Eff.
85.0
DeepSeek-V4-Flash (思考 · API)(Thinking · API)
🧠 MoE · 284B (13B active) ☁️ OpenCode API 📅 2026-06-19
ToolCall-15
100
BugFind-15
93.3
HermesAgent-20
90
🔄 上限分 94.0 → 重试 -9 → 实用 85.0🔄 Max 94.0 → retries -9 → Eff. 85.0
2026-06-19 测试tested · ⚡ 作为参考reference 查看详情 →View Details →
2
上限Max
93.5
实用Eff.
75.5

Ornith-1.0-35B-MTP-APEX (思考 · MTP)(Thinking · MTP)

Ornith-1.0-35B-MTP-APEX-I-Compact.gguf · Q4_K_M · 15.85 GB
🧠 MoE (3B active) 📦 15.85 GB 🎮 RTX 5070 Ti 🌡️ temp=0.6 top_p=0.95 top_k=20
ToolCall-15
100
BugFind-15
93
HermesAgent-20
89
🔄 上限分 93.5 → 重试 -18 → 实用 75.5🔄 Max 93.5 → retries -18 → Eff. 75.5
2026-06-26 测试tested 查看详情 →View Details → HF
1
上限Max
95.0
实用Eff.
90.0

Ornith-1.0-35B-Heretic-MTP-APEX (思考 · Heretic · MTP)(Thinking · Heretic · MTP)

Ornith-1.0-35B-Heretic-MTP-APEX-I-Compact.gguf · Q4_K_M · 15.85 GB
🧠 MoE (3B active) 📦 15.85 GB 🎮 RTX 5070 Ti 🔥 Heretic Abliterated 🌡️ temp=0.6 top_p=0.95 top_k=20
ToolCall-15
100
BugFind-15
98
HermesAgent-20
89
🔄 上限分 95.0 → 重试 -5 → 实用 90.0🔄 Max 95.0 → retries -5 → Eff. 90.0
2026-07-12 测试tested 查看详情 →View Details → HF
-
上限Max
92.2
实用Eff.
76.2

Ornith-1.0-35B-Heretic-MTP-APEX (无思考 · Heretic · MTP)(No Thinking · Heretic · MTP)

Ornith-1.0-35B-Heretic-MTP-APEX-I-Compact.gguf · Q4_K_M · 15.85 GB
🧠 MoE (3B active) 📦 15.85 GB 🎮 RTX 5070 Ti 🔥 Heretic Abliterated 🌡️ temp=0.6 top_p=0.95 top_k=20
ToolCall-15
97
BugFind-15
97
HermesAgent-20
85
🔄 上限分 92.2 → 重试 -16 → 实用 76.2🔄 Max 92.2 → retries -16 → Eff. 76.2
2026-07-12 测试tested 查看详情 →View Details → HF
2
上限Max
93.2
实用Eff.
85.2

Ornith-1.0-35B-MTP-APEX (无思考 · MTP)(No Thinking · MTP)

Ornith-1.0-35B-MTP-APEX-I-Compact.gguf · Q4_K_M · 15.85 GB
🧠 MoE (3B active) 📦 15.85 GB 🎮 RTX 5070 Ti 🌡️ temp=0.6 top_p=0.95 top_k=20
ToolCall-15
100
BugFind-15
92
HermesAgent-20
89
🔄 上限分 93.2 → 重试 -8 → 实用 85.2🔄 Max 93.2 → retries -8 → Eff. 85.2
2026-06-27 测试tested 查看详情 →View Details → HF
-
上限
91.2
实用
71.2

Agents-A1 (思考 · MTP)(Thinking · MTP)

Agents-A1-APEX-I-Compact.gguf · Q4_K_M · 16.1 GB
🧠 MoE (3B active) 📦 16.1 GB 🎮 RTX 5070 Ti 16GB + 128GB RAM ⚡ In 804 t/s · Out 55.6 t/s 🌡️ temp=0.85 top_p=0.95 top_k=20 min_p=0.0 presence=1.1 rep=1.0
ToolCall-15
100
BugFind-15
88
HermesAgent-20
87
🔄 上限分 91.2 → 重试 -20 → 实用 71.2🔄 Max 91.2 → retries -20 → Eff. 71.2
2026-07-03 测试tested 查看详情 →View Details → HF
-
上限
93.1
实用
57.1

Agents-A1 (无思考 · MTP)(No Think · MTP)

Agents-A1-APEX-I-Compact.gguf · Q4_K_M · 16.1 GB
🧠 MoE (3B active) 📦 16.1 GB 🎮 RTX 5070 Ti 16GB + 128GB RAM ⚡ In 804 t/s · Out 55.6 t/s 🌡️ temp=0.85 top_p=0.95 top_k=20 min_p=0.0 presence=1.1 rep=1.0
ToolCall-15
97
BugFind-15
100
HermesAgent-20
85
🔄 上限分 93.1 → 重试 -36 → 实用 57.1🔄 Max 93.1 → retries -36 → Eff. 57.1
2026-07-03 测试tested 查看详情 →View Details → HF
-
上限Max
85.6
实用Eff.
81.6

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 🌡️ temp=1.0 top_p=0.95 top_k=20 rep=1.5
ToolCall-15
93
BugFind-15
87
HermesAgent-20
79
🔄 上限分 85.6 → 重试 -4 → 实用 81.6🔄 Max 85.6 → retries -4 → Eff. 81.6
2026-07-07 测试tested 查看详情 →View Details → HF
-
上限Max
75.0
实用Eff.
75.0

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 🌡️ temp=0.7 top_p=0.8 top_k=20 rep=1.5
ToolCall-15
83
BugFind-15
83
HermesAgent-20
63
🔄 上限分 75.0 → 重试 -0 → 实用 75.0🔄 Max 75.0 → retries -0 → Eff. 75.0
2026-07-08 测试tested 查看详情 →View Details → HF
-
上限Max
88.5
实用Eff.
87.5

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 🌡️ temp=1.0 top_p=0.95 top_k=20 rep=1.5
ToolCall-15
100
BugFind-15
87
HermesAgent-20
81
🔄 上限分 88.5 → 重试 -1 → 实用 87.5🔄 Max 88.5 → retries -1 → Eff. 87.5
2026-07-08 测试tested 查看详情 →View Details → HF

⚡ 关于 BenchLocalAbout BenchLocal

BenchLocal 是一个本地优先的桌面应用，用于运行、比较和管理可安装的 LLM Bench Packs。支持本地或远程模型，通过可安装的测试包对模型进行标准化评估。BenchLocal is a local-first desktop app for running, comparing, and managing installable LLM Bench Packs. Supports local or remote models with standardized evaluation via installable test packs.

📋 测试项目Test Suites

  • ToolCall-15 — 15 题工具调用测试，覆盖参数提取、多轮上下文、并行调用等ToolCall-15 — 15-question tool calling test covering parameter extraction, multi-turn context, parallel calls
  • BugFind-15 — 15 题跨语言代码调试，含 Trap 陷阱题，难度 Easy~ExpertBugFind-15 — 15-question cross-language debugging with Trap questions, Easy~Expert
  • HermesAgent-20 — 20 题 Agent 场景测试，覆盖记忆管理、技能创建、调度投递等HermesAgent-20 — 20-question Agent scenarios covering memory management, skill creation, scheduling
  • 能力上限 — 加权原始分 = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4（模型最高能拿多少分）Max Score — Weighted raw score = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4 (highest possible score)
  • 实用得分 — 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）Effective Score — Max Score − retry penalty (first pass = no penalty, each retry = +1pt, failed retries excluded)
  • 能力上限排序（看天花板）、实用得分排序（看实际表现）、三项专项排序Sort by max score (ceiling), effective score (actual performance), or per-suite ranking

📊 测试总结Test Summary

基于 BenchLocal 在同一硬件环境下的本地推理测试，横向对比 19 个模型配置Based on BenchLocal local inference tests on the same hardware, comparing 19 model configurations

🏆 核心发现Key Findings

  • 能力上限 vs 实用得分差距显著 — 重试次数越多，两者的差距越大：N2-Mini 差 27 分（87.2→60.2），27B 无思考版仅差 3 分（87.4→84.4）Big gap between Max Score and Effective Score — The more retries, the bigger the gap: N2-Mini drops 27 pts (87.2→60.2), while 27B no-think only drops 3 (87.4→84.4)
  • ToolCall 满分常见ToolCall perfect scores are common
  • DeepSeek-V4-Flash（284B MoE）拿到 94.0 分，实用 85.0 分，三项均衡DeepSeek-V4-Flash (284B MoE) scored 94.0, effective 85.0, well-balanced across all three
  • BugFind BF-03 Trap 题全场最难 — 多数模型在"代码没bug"场景翻车，无法判断代码是正确的BugFind BF-03 Trap was the hardest — Most models fail in "code has no bug" scenarios, unable to judge that code is correct

⚠️ 已知局限Known Limitations

  • BF-03（代码没bug）和 BF-10（红鲱鱼）是多数模型的软肋BF-03 (code has no bug) and BF-10 (red herring) are weak points for most models
  • HA-07（代码批量处理）大部分失败 — execute_code 能力普遍薄弱HA-07 (batch code processing) mostly failed — execute_code capability is generally weak
  • HA-16（消息投递）全军覆没 — 所有模型都无法找到正确的消息通道HA-16 (message delivery) completely failed — No model could find the correct message channel
  • 重试成本差异巨大Retry costs vary dramatically

💡 测试环境Test Environment

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