Update Note: This is a follow-up to my previous article “Finding the Perfect Partner for Your AI Agents - oh-my-opencode Model Selection Guide”. Back then, I configured the system based on 5 models: GLM-4.7, GLM-5, MiniMax-M2.5, DeepSeek-V3.2, and Kimi-K2.5. Recently, I added 4 Qwen series models (qwen3.5-plus, qwen3-max, qwen3-coder-next, qwen3-coder-plus), prompting a fresh re-evaluation of the optimal model for each Agent.

Why Re-evaluate?

After publishing the previous article, my model subscription changed: Alibaba Cloud Bailian added Qwen series models. The benchmark data for these 4 new models is impressive:

• qwen3-max: GPQA 86.1% (scientific reasoning ceiling), LiveCodeBench 91.4% (coding reasoning leader)
• qwen3-coder-next: SWE-bench 70.6% + $0.12/M (value champion)
• qwen3.5-plus: 1M context + multimodal (long document powerhouse)

This data compelled me to re-examine my previous configuration. After deep research and benchmark analysis, I arrived at this new configuration scheme.

Why “Role-Based Allocation”?

oh-my-opencode has an interesting architecture that splits workflows into specialized Agents:

• Sisyphus: The conductor, orchestrates tasks and delegates work
• Hephaestus: Deep worker, executes tasks end-to-end
• Oracle: Complex debugging, architecture consultant
• Librarian: Documentation retrieval, external library queries
• Explore: Codebase search specialist
• Metis: Pre-planning consultant, identifies implicit intents
• Momus: Plan reviewer
• Prometheus: Strategic planner
• Multimodal-looker: Image/video analysis
• Atlas: Main model for UI interactions

Each Agent has different responsibilities and requires different model capabilities. Just like in a team—some excel at design, others at coding, others at documentation—models should be assigned accordingly.

What Changed This Time?

Compared to the previous configuration, here are the core changes:

In short: Reasoning scenarios upgrade to qwen3-max, coding scenarios switch to qwen3-coder-next, long document scenarios use qwen3.5-plus, while three specialists remain unchanged.

My Complete Model Lineup

Before diving into allocation, let me show you all the “players” I have:

Key Discoveries: The New Contenders

qwen3-max: The New “Reasoning King”

This was my biggest discovery from this research:

• GPQA 86.1% — Strongest scientific reasoning in publicly available data
• LiveCodeBench v6 91.4% — Coding reasoning ceiling
• Test-time Scaling + Early Stop Detection — Automatically determines when to stop thinking, no wasted compute

What does this mean? If you need deep reasoning, architecture analysis, or complex debugging, qwen3-max is currently the best choice.

qwen3-coder-next: The Value Champion

• SWE-bench Verified 70.6% — Close to MiniMax-M2.5’s 80.2%
• Output speed 151.5 tokens/sec — 2nd in its class
• Response time 11.68 seconds — vs MiniMax-M2.5’s 43.03 seconds
• Price $0.12/M — Cheaper than any competitor

What does this mean? If you need high-frequency calls and fast responses (code exploration, quick fixes), qwen3-coder-next is the obvious choice.

qwen3.5-plus: Multimodal + Long Context New Option

• 1M context — Largest context window currently available
• Native multimodal — Supports text, images, video
• Apache 2.0 open source — Self-deployable
• Competitive pricing — $0.12-0.26/M input

What does this mean? For scenarios requiring long documents or multimodal content (document retrieval, UI interaction, writing), qwen3.5-plus is an ideal choice.

Core Configuration Strategy

After careful consideration, I established these configuration principles:

1. The Conductor Needs Low Hallucination

Sisyphus → GLM-5

Why not qwen3-max? Because as the conductor, reliability trumps reasoning depth. GLM-5’s hallucination rate is 56% lower than GLM-4.7, making it more reliable for task orchestration.

2. Deep Reasoning Scenarios Use the Strongest Reasoning

Prometheus, Oracle, Ultrabrain → qwen3-max

These scenarios require deep reasoning: strategic planning, architecture consulting, complex logic analysis. qwen3-max’s GPQA 86.1% and LiveCodeBench 91.4% are currently the ceiling.

3. High-frequency Coding Scenarios Use Speed Models

Explore, Quick, Deep, Hephaestus → qwen3-coder-next

These scenarios have high call frequency and need fast responses. qwen3-coder-next’s 151.5 tokens/sec and $0.12/M price make it the best choice.

4. Multimodal and Long Context Use the All-rounder

Librarian, Metis, Atlas, Visual-engineering, Artistry, Writing → qwen3.5-plus

These scenarios require processing long documents or multimodal content. qwen3.5-plus’s 1M context and native multimodal support are key.

5. Special Scenarios Retain Specialists

Momus → MiniMax-M2.5 (SWE-bench 80.2% highest score, plan review needs coding accuracy)

Multimodal-looker → Kimi-K2.5 (Video understanding capability, qwen3.5-plus doesn’t support video yet)

Final Configuration

Agents Configuration

Categories Configuration

Configuration Code

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

{
  "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-opencode/dev/assets/oh-my-opencode.schema.json",
  "agents": {
    "hephaestus": {
      "model": "bailian-coding-plan/qwen3-coder-next"
    },
    "oracle": {
      "model": "bailian-coding-plan/qwen3-max-2026-01-23"
    },
    "librarian": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    },
    "explore": {
      "model": "bailian-coding-plan/qwen3-coder-next"
    },
    "multimodal-looker": {
      "model": "volcengine-coding/kimi-k2.5"
    },
    "prometheus": {
      "model": "bailian-coding-plan/qwen3-max-2026-01-23"
    },
    "metis": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    },
    "momus": {
      "model": "volcengine-coding/minimax-m2.5"
    },
    "atlas": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    },
    "sisyphus": {
      "model": "bailian-coding-plan/glm-5"
    }
  },
  "categories": {
    "visual-engineering": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    },
    "ultrabrain": {
      "model": "bailian-coding-plan/qwen3-max-2026-01-23"
    },
    "deep": {
      "model": "bailian-coding-plan/qwen3-coder-next"
    },
    "artistry": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    },
    "quick": {
      "model": "bailian-coding-plan/qwen3-coder-next"
    },
    "unspecified-low": {
      "model": "bailian-coding-plan/qwen3-coder-next"
    },
    "unspecified-high": {
      "model": "bailian-coding-plan/qwen3-max-2026-01-23"
    },
    "writing": {
      "model": "bailian-coding-plan/qwen3.5-plus"
    }
  }
}

Model Distribution Visualization

1
2
3
4
5
6

qwen3-coder-next  ████████████ 5 positions (hephaestus, explore, deep, quick, unspecified-low)
qwen3.5-plus      ██████████████ 6 positions (librarian, metis, atlas, visual-engineering, artistry, writing)
qwen3-max         ██████████ 4 positions (prometheus, oracle, ultrabrain, unspecified-high)
GLM-5             ██ 1 position (sisyphus)
MiniMax-M2.5      ██ 1 position (momus)
Kimi-K2.5         ██ 1 position (multimodal-looker)

Lessons Learned

1. Don’t Blindly Chase “Latest and Greatest”

qwen3-max’s GPQA 86.1% is impressive, but GLM-5’s low hallucination characteristic is more important for orchestration scenarios. Choose based on actual needs, not just benchmark scores.

2. Optimize High-frequency Scenarios Separately

Explore and Quick are high-frequency Agents. Using expensive models would quickly deplete quotas. qwen3-coder-next’s $0.12/M eliminates this concern entirely.

3. Special Capabilities Need Specialists

Video understanding is currently only supported by Kimi-K2.5, and plan review needs MiniMax-M2.5’s SWE-bench highest score. These special scenarios can’t be replaced by “general-purpose strong models.”

4. Testing Beats Theory

After configuration, I recommend testing these typical scenarios:

• Code search (triggers Explore)
• Document retrieval (triggers Librarian)
• Visual analysis (triggers Multimodal-looker)
• Complex architecture design (triggers Oracle or Ultrabrain)

Conclusion

This adjustment is a comprehensive upgrade to my previous configuration. There’s no best model, only the most suitable model—this still holds true, but as new models are added, the “most suitable” answer changes.

After completing this configuration, the system’s efficiency and effectiveness improved noticeably. Each Agent is doing what it excels at, and the collaboration is smoother.

If you’re also using oh-my-opencode, I recommend adjusting your configuration based on your use cases and available models. After all, finding the right partner is what doubles your productivity.