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📅 2026-07-03🌡️ temp=0.85 top_p=0.95 top_k=20 min_p=0.0 presence=1.1 rep=1.0
能力上限Max Score
91.2
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
71.2
−20 重试扣分−20 retry penalty
ToolCall-15
100
15/15 通过 · 100%passed · 100%
BugFind-15
88
13/15 通过 · 87%passed · 87%
🔄 重试 6 次🔄 6 retries
HermesAgent-20
87
15/20 通过 · 75%passed · 75%
🔄 重试 9 次🔄 9 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.

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 91.2（排名第 4），实用得分 71.2（扣分 20 分，主要来自 TC-03 的 3 次重试、HA-02/05/08 的多次重试和 HA-07/10/16/17/20 的反复失败）。This model: Max Score 91.2 (Rank #4), Effective Score 71.2 (−20 penalty, mainly from TC-03's 3 retries, HA-02/05/08 retries and HA-07/10/16/17/20 repeated failures).

ToolCall-15 (15/15)ToolCall-15 (15/15)
BugFind-15 (13/15)BugFind-15 (13/15)
HermesAgent-20 (15/20)HermesAgent-20 (15/20)
❌ 失败 (7)❌ Failed (7)
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

Agents-A1（思考模式）展现了强劲的基础能力——ToolCall-15 满分通过（100），这是其 agentic coding 训练的直接体现。BugFind-15 能力上限 88 分，仅 BF-03（语法陷阱）和 BF-10（红鲱鱼题）失败。HermesAgent-20 能力上限 87 分，记忆和技能维度表现优秀，但 Delegation/Recovery/Boundaries 维度仅 73 分。Agents-A1 (Thinking) shows strong fundamentals — perfect ToolCall-15 (100), reflecting its agentic coding training. BugFind-15 ceiling at 88, failing only BF-03 (syntax trap) and BF-10 (red herring). HermesAgent-20 ceiling at 87, excelling in Memory and Skills but weak in Delegation/Recovery/Boundaries (73).

稳定性是最大短板：HermesAgent-20 有 5 题始终未通过（HA-07/10/16/17/20），各尝试 8 次仍失败——这些场景涉及并行委派、模糊破坏性请求等复杂 agent 行为。BugFind-10 是红鲱鱼陷阱，模型过度思考反而修坏了正确代码。总计 20 次重试扣了 20 分。Stability is the biggest weakness: HermesAgent-20 has 5 permanently failed scenarios (HA-07/10/16/17/20), each with 8 attempts — these involve parallel delegation, ambiguous destructive requests. BF-10 is a red herring where overthinking led to modifying correct code. Total 20 retries cost 20 points.

能力天花板 91.2 与 Ornith（93.5）和 DeepSeek V4（94.0）接近，但 20 次重试让实用分只有 71.2——agentic 场景的稳定性仍是核心瓶颈。作为 InternScience 基于 Qwen3.5-35B-A3B 微调的 agentic 模型，35B 参数量（3B 激活）在工具调用上达到了满分水平，但在复杂 agent 行为（委派、边界、恢复）上还有提升空间。Ceiling of 91.2 is close to Ornith (93.5) and DeepSeek V4 (94.0), but 20 retries drag effective score to 71.2. As InternScience's agentic model fine-tuned from Qwen3.5-35B-A3B, it achieves perfect tool calling but has room to improve in complex agent behaviors.