Ornith-1.0-35B-Heretic-MTP-APEX (无思考 · Heretic · MTP)(No Thinking · Heretic · 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
92.2
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
76.2
−16 重试扣分−16 retry penalty
ToolCall-15
97
14/15 通过 · 93%passed · 93%
BugFind-15
97
14/15 通过 · 93%passed · 93%
🔄 重试 8 次🔄 8 retries
HermesAgent-20
85
14/20 通过 · 70%passed · 70%
🔄 重试 7 次🔄 7 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 92.2，实用得分 76.2（扣 16 分）。相比 Ornith-1.0-35B-Heretic-MTP-APEX 思考版（95.0 / 90.0）差距明显：天差地别的不是能力，是稳定性。heretic 在思考模式下是"助力"，但关闭思考后变成"裸奔"——重试多了 11 次却救不回原本就达不到的能力短板。This model: Max Score 92.2, Effective Score 76.2 (−16 penalty). Compared to Ornith-1.0-35B-Heretic-MTP-APEX Thinking (95.0 / 90.0) the gap is huge — capability barely drops (−2.8), but stability worsens by +11 retries. heretic ablation is an asset with thinking, but disabling thinking leaves the model "naked": more retries can't compensate for absent deliberation.

ToolCall-15 (14/15)ToolCall-15 (14/15)
BugFind-15 (14/15)BugFind-15 (14/15)
HermesAgent-20 (14/20)HermesAgent-20 (14/20)
❌ 失败 (8)❌ Failed (8)
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-Heretic-MTP（无思考模式）展现了与思考版截然不同的画像：能力上限 -2.8（95.0 → 92.2），但实用得分暴跌 -13.8（90.0 → 76.2）。重试从 5 次激增到 16 次，每个失败的 HA 题都跑满 6 次尝试仍未通过。这与原版（无 heretic）的"无思考更稳"结论完全相反。Ornith-1.0-35B-Heretic-MTP (No Thinking) shows a very different profile vs Thinking: ceiling drops −2.8 (95.0 → 92.2), but effective score collapses −13.8 (90.0 → 76.2). Retries jump from 5 to 16, with every failed HA scenario running all 6 attempts. This fully reverses the original (non-heretic) conclusion that "no-thinking is more reliable".

失败项目集中在 HermesAgent：HA-06/07/16/17/20 共 5 道大题，每次都是 6 次重试全军覆没。同时 BF-09 重试 4 次、BF-10 重试 2 次、B>>>F-04/15 各重试 1 次——重试代价大幅增加。heretic 在思考型模型里是"去除过度分析的助力"，但关闭思考后模型"裸奔"，最终被卡在需要谨慎推理的场景上（多步骤流程、并行委派、跨平台消息路由）。Failures cluster in HermesAgent: HA-06/07/16/17/20 (5 scenarios × 6 attempts each, all failed). Plus BF-09 (4 retries), BF-10 (2), BF-04/15 (1 each) — retry cost skyrockets. With thinking, heretic ablation removes "over-analysis". Without thinking the model is "running naked" and gets stuck on scenarios that demand deliberation: multi-step flows, parallel delegation, cross-platform message routing.

能力天花板 92.2，实用 76.2 扣 16 分——heretic 模型请开思考。无思考 + heretic 是反模式。Ceiling 92.2, effective 76.2 with −16 penalty — always enable thinking on heretic models. No-thinking + heretic is an anti-pattern.