Ornith-1.0-35B-Heretic-MTP-APEX (思考 · Heretic · MTP)(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
95.0
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
90.0
−5 重试扣分−5 retry penalty
ToolCall-15
100
15/15 通过 · 100%passed · 100%
BugFind-15
98
14/15 通过 · 93%passed · 93%
🔄 重试 2 次🔄 2 retries
HermesAgent-20
89
16/20 通过 · 80%passed · 80%
🔄 重试 3 次🔄 3 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 95.0（全场第 1），实用得分 90.0（仅扣 5 分）。相比原版 Ornith-1.0-35B-MTP-APEX（思考），能力天花板提升 +1.5，实用得分提升 +14.5。heretic abliteration 让 BF-10 红鲱鱼陷阱（0 → 100）和稳定性都显著改善。This model: Max Score 95.0 (Rank #1 overall), Effective Score 90.0 (−5 penalty). Compared to the original Ornith-1.0-35B-MTP-APEX (Thinking), the ceiling rises +1.5 and the effective score jumps +14.5. heretic abliteration dramatically improves the BF-10 red-herring trap (0 → 100) and overall stability.

ToolCall-15 (15/15)ToolCall-15 (15/15)
BugFind-15 (15/15)BugFind-15 (15/15)
HermesAgent-20 (16/20)HermesAgent-20 (16/20)
❌ 失败 (5)❌ Failed (5)
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（思考模式）展现了与原版不同的稳定性：ToolCall-15 满分零重试，BugFind-15 能力上限 98 分（仅 BF-03 陷阱题掉到 60），HermesAgent-20 在记忆/工作区/技能三个维度仍然满分。heretic abliteration 让模型在 BF-10 红鲱鱼陷阱题上从原版的 0 分跃升到 100 分——这是核心改进：模型不再"过度分析"，对正确代码不再强行重写。Ornith-1.0-35B-Heretic-MTP (Thinking) shows a different stability profile: TC perfect with zero retries, BF ceiling 98 (only the BF-03 trap question drops to 60), HA still perfect on Memory/Workspace/Skills. The heretic ablation is the key improvement — model no longer "overthinks" the BF-10 red herring (0 → 100), leaving correct code untouched.

稳定性大幅改善：BF-13 从原来的 5 次重试降到 0 次重试一次通过；总重试从 18 次降到 5 次，是全场思考型模型中最低的。仍存在的弱点是 HermesAgent 的委派/恢复/边界维度 60 分（HA-16/17/20 均 4 次重试仍失败），与 BF-03 Rust trap 是 4 个唯一未攻克的难题。Stability improved dramatically: BF-13 dropped from 5 retries to 0 retries (clean first pass). Total retries dropped from 18 to 5 — the lowest among all Thinking-mode models. The remaining weakness is HermesAgent Delegation/Recovery/Boundaries at 60 (HA-16/17/20 all failed despite 4 attempts), plus BF-03 Rust trap — these 4 unsolved problems mark the frontier.

能力天花板 95.0 全场第一，实用得分 90.0 仅扣 5 分——heretic 把"想太多"的双刃剑变成了稳定战斗力。Ceiling of 95.0 ranks #1 overall, effective 90.0 with only −5 penalty — heretic turned the "overthinking" double-edged sword into reliable combat power.