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 16GB + 128GB RAM📅 2026-06-26🌡️ temp=0.6 top_p=0.95 top_k=20
能力上限Max Score
93.5
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
75.5
−18 重试扣分−18 retry penalty
ToolCall-15
100
15/15 通过 · 100%passed · 100%
BugFind-15
93
14/15 通过 · 93%passed · 93%
🔄 重试 12 次🔄 12 retries
HermesAgent-20
89
15/20 通过 · 75%passed · 75%
🔄 重试 6 次🔄 6 retries

📐 计分规则Scoring Rules

能力上限 = 加权原始分 = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4。反映模型的最高能力水平。Max Score = Weighted raw score = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4. Reflects the model's highest capability level.

实用得分 = 能力上限 − 重试扣分。每题首次通过不扣分；重试后才通过的题目每重试1次扣1分；重试后仍失败的不参与扣分。Effective Score = Max Score − retry penalty. First pass = no penalty; each retry before passing = −1pt; failed retries excluded.

本模型能力上限 93.5（排名第 2），实用得分 75.5（扣分 18 分，主要来自 BF-13 的 5 次重试和 HA-02/HA-16/HA-17/HA-20 的高重试）。This model: Max Score 93.5 (Rank #2), Effective Score 75.5 (−18 penalty, mainly from BF-13's 5 retries and HA-02/HA-16/HA-17/HA-20 high retries).

ToolCall-15 (15/15)ToolCall-15 (15/15)
BugFind-15 (14/15)BugFind-15 (14/15)
HermesAgent-20 (15/20)HermesAgent-20 (15/20)
❌ 失败 (6)❌ Failed (6)
题号ID分数Score结果Result重试Retry说明Summary
题号ID分数Score结果Result重试Retry说明Summary
题号ID分数Score结果Result重试Retry说明Summary
题号ID分数Score结果Result重试Retry失败原因Failure Reason

📊 模型评估总结Model Evaluation Summary

Ornith-1.0-35B-MTP（思考模式）展现了扎实的基础能力——ToolCall-15 满分通过，BugFind-15 能力上限 93 分，HermesAgent-20 记忆和技能维度均满分。但"思考过多"导致它在红鲱鱼陷阱题（BF-10）上两次翻车，强行修了本来正确的代码。Ornith-1.0-35B-MTP (Thinking) demonstrates solid foundational capabilities — perfect ToolCall-15 score, 93 on BugFind-15 ceiling, and full marks on Memory and Skills in HermesAgent-20. However, "overthinking" caused it to fail the red herring trap (BF-10) twice, "fixing" code that was already correct.

稳定性是最大瓶颈：BugFind-13 花了 5 次重试才通过（连接重置问题），HermesAgent 的委派/恢复/边界维度仅 68 分（HA-16/17/20 均高重试仍失败）。总重试 18 次扣了 18 分，把 93.5 的能力天花板拉到了 75.5 的实用分。Stability is the biggest bottleneck: BF-13 took 5 retries (connection resets), HermesAgent's Delegation/Recovery/Boundaries scored only 68 (HA-16/17/20 all failed despite high retries). Total 18 retries cost 18 points, dropping from 93.5 ceiling to 75.5 effective.

能力天花板 93.5 全场第二，但 18 次重试让实用分只有 75.5——思考型模型的"想太多"双刃剑。Ceiling of 93.5 ranks #2, but 18 retries drag effective score to 75.5 — the double-edged sword of overthinking.