9Router Combo Cheatsheet for Daily Use

Why do I need a combo cheatsheet?#

As the number of models grows, the question is no longer “which model is the most powerful” but which model fits best for each task.

With 9Router, I don’t want to sit and remember every time I work:

• which combo is good for fast coding
• which combo is good for deep thinking
• which combo is good for review
• which combo is good for reading images or screenshots

This is a practical note to open and pick a combo in seconds.

How I categorize combos#

1. Separate by type of work, not by model name
2. Within each combo, select models by role
3. Write short notes on each model’s strength for easy recall
4. Start with lightweight combos, only upgrade when the problem gets “intense”

Quick selection map#

1. coding-light#

Use when

• daily coding
• CRUD
• fixing medium bugs
• quick scaffolding
• light refactoring
• agent editing code continuously

Personality of this combo

This is a coder-first combo. The goal is to write code quickly and stably, without pulling in reasoning-heavy models for tasks that don’t need that “heavy lifting”.

Models#

• qw/qwen3-coder-plus — main coder, good balance between quality and speed
• kr/qwen3-coder-next — good for agent coding, strong in multi-step flows
• if/qwen3-coder-plus — backup coder from the same thinking family as Qwen coder
• qw/qwen3-coder-flash — fastest in the group, good for small patches and short tasks

Mantra

Normal coding, no philosophy needed, just coding-light.

2. coding-hard#

Use when

• hard features
• multi-file refactoring
• tricky logic bugs
• tasks needing both coding and thinking
• agent must plan before fixing

Personality of this combo

This is engineering mode combo. It doesn’t just write code; it has enough brain to think before producing patches.

Models#

• qw/qwen3-coder-plus — main code writer
• kr/claude-sonnet-4.5 — planner/reviewer, strong at structure and explanation
• if/deepseek-v3.2 — auxiliary reasoning, connects logic well
• kr/qwen3-coder-next — supplementary coder, good for extended multi-step tasks

Mantra

Hard coding, many steps, need to think while doing → coding-hard.

3. reasoning#

Use when

• finding root causes
• very hard debugging
• analyzing long logs
• debating solutions
• complex algorithms or logic flows
• not ready to code yet

Personality of this combo

This combo is for when you need to say:

Wait, don’t fix yet. Sit down and think first.

Models#

• if/deepseek-r1 — digs deep into logic, great for finding root causes
• if/kimi-k2 — holds long context, synthesizes lots of information
• if/qwen3-max — balanced reasoning, versatile, easy to delegate
• if/qwen3-235b-a22b-thinking-2507 — specialized deep thinker, good for critique and persistent thinking

Mantra

When the problem’s cause isn’t clear, don’t touch code, use reasoning.

4. architect#

Use when

• writing specs
• proposing architecture
• dividing modules or services
• selecting patterns
• writing docs, proposals, design notes
• comparing approaches

Personality of this combo

This combo is for design and expression. It’s not aimed at producing quick patches, but at making good decisions and writing readable documentation.

Models#

• kr/claude-sonnet-4.5 — writes fluently, beautiful structure, great for design/spec
• if/kimi-k2 — holds long context, good at synthesizing documents
• if/qwen3-max — balanced between thinking and practicality
• if/deepseek-r1 — logic critique, spots architecture holes

Mantra

Writing spec, design, proposal, or architecture notes → architect.

5. eazy#

Use when

• quick questions
• text transformation
• fixing short snippets
• light brainstorming
• casual chat
• many small tasks

Personality of this combo

This is the fast lane combo. Not every task needs calling out the big brain council. Many tasks just need a fast, compact, smart-enough combo to get done.

Models#

• gc/gemini-3-flash-preview — very fast, good for Q&A and daily tasks
• kr/claude-haiku-4.5 — compact, fast, stable for text and general chat
• if/deepseek-v3.2 — fast but still has brain, good for slightly technical tasks
• qw/qwen3-coder-flash — fast for code snippets or small patches

Mantra

Trivial tasks, need speed, no overkill → eazy.

6. vision#

Use when

• reading error screenshots
• analyzing UI
• viewing diagrams
• images with text
• looking at interfaces to infer flow or code

Personality of this combo

This combo is for everything with visual elements. When there are screenshots, UI, diagrams, or error photos, using the right vision combo helps models stop guessing from thin air.

Models#

• if/qwen3-vl-plus — main vision model, good for images and technical analysis
• gc/gemini-3-pro-preview — strong multimodal, good for image + reasoning
• qw/vision-model — vision fallback, adds a backup perspective

Mantra

Got images? Don’t force text-only combos to guess. Just use vision.

7. review#

Use when

• reviewing PRs
• auditing code
• critiquing solutions
• checking edge cases
• spotting potential bugs
• verifying output from other combos

Personality of this combo

This combo is for inspection. Not for creating the first solution, but for checking if that solution has anything suspicious.

Models#

• kr/claude-sonnet-4.5 — fluent review, clear comments, easy to read
• if/deepseek-r1 — deep logic inspection, good for finding hidden bugs
• if/qwen3-max — balanced reviewer, practical

Mantra

Already have a solution, now need to inspect it → review.

Practical workflows#

Case 1: Building a backend feature#

Start with coding-light.

If the task starts showing these signs:

• many files
• complex flow
• agent needs to plan first
• fixing one spot pulls three other spots

then upgrade to coding-hard.

Case 2: Hard bug#

Start with reasoning to find the root cause.

Once the fix direction is clear, switch to coding-hard to produce the patch. This approach is often much better than “fixing by feeling”.

Case 3: Writing architecture documentation#

Use architect to establish the approach, select patterns, and write specs.

When the first draft is done, throw it to review for self-critique.

Case 4: Got a screenshot or image#

Start with vision. After understanding the image or UI flow, if code fixes are needed, switch to coding-light or coding-hard.

Case 5: Have code but not fully confident#

Use review. This is a valuable step when you want to check edge cases, code smells, or risks before merging.

Pocket note#

coding-light  = daily coding
coding-hard   = hard coding, many steps
reasoning     = deep thinking, root cause
architect     = spec, design, proposal
eazy          = quick questions, small tasks
vision        = images, UI, screenshots
review        = inspect solutions / PRs / edge cases

Conclusion#

The key isn’t which model is the most powerful, but putting the right model in the right role.

This cheatsheet helps me reduce hesitation time when picking combos:

• don’t mistakenly call a thinker for quick coding
• don’t rush to throw coders at root cause problems
• don’t use casual chat combos for review
• don’t make text-only models guess from thin air when there are images

If you’re also using multiple combos in 9Router, treat this as a mini playbook: open it, pick the right lane, then get to work.