LM Studio Companion Guide

LM Studio Companion: Walkthrough, Prompts, and Benchmarks

Use this companion guide after you finish the local setup guide.

1. Screenshot Walkthrough Checklist

Use this checklist while opening LM Studio for the first time.

1. Home screen: confirm LM Studio opens without errors.
2. Model search: find a model like Qwen2.5 3B Instruct.
3. Download panel: choose a quantization, prefer Q4_K_M.
4. Model loaded: confirm memory usage looks stable after load.
5. Chat tab: send a test prompt and verify response speed.
6. Settings tab: set context window to 4096 first.
7. GPU settings: enable maximum stable GPU offload.
8. API server (optional): enable local OpenAI-compatible endpoint.
9. Final validation: restart LM Studio and re-run one test prompt.

Tip for junior developers: capture one screenshot per step and save with names like:

• 01-home.png
• 02-model-search.png
• 03-chat-test.png

This makes troubleshooting and team handover much easier.

2. Coding Prompt Starter Pack

Copy and paste these prompts directly into LM Studio.

Prompt 1: Explain code simply

You are a senior developer mentoring a junior developer.
Explain this code in plain language.
Then list:
1) what is good
2) what is risky
3) one safe improvement

Code:
[paste code here]

Prompt 2: Debug a failing function

Help me debug this function.
Return your answer in this format:
- Suspected root cause
- How to reproduce
- Minimal fix
- Final corrected code

Function:
[paste function]

Error:
[paste error message]

Prompt 3: Refactor without behavior change

Refactor this code for readability.
Rules:
- Do not change behavior
- Keep function names
- Add short comments only where logic is not obvious
- Return complete updated code

Code:
[paste code]

Prompt 4: Generate tests first

Write tests for this code.
Use clear test names and include edge cases.
If assumptions are needed, state them first.

Code:
[paste code]

Prompt 5: API integration helper

I am building a Next.js API route.
Given this requirement and input/output shape, generate:
1) route handler
2) validation
3) error handling
4) one example request and response

Requirement:
[paste requirement]

3. Mini Benchmark Checklist (15 minutes)

Run this when choosing between two or three local models.

A. Prepare

1. Pick 3 prompts you will reuse for every model:
   • one coding prompt
   • one debugging prompt
   • one explanation prompt
2. Keep settings fixed across models:
   • same context window
   • same temperature
   • same machine state (close heavy apps)

B. Measure

For each model, record:

• Time to first token (seconds)
• Approximate tokens per second
• Memory pressure (none, medium, high)
• Response quality (1 to 5)
• Hallucination rate (low, medium, high)

Use a simple score:

$$
\text{Model Score} = 0.35Q + 0.30S + 0.20R + 0.15A
$$

Where:

• $Q$ = quality score (1 to 5)
• $S$ = speed score (1 to 5)
• $R$ = reliability score (1 to 5)
• $A$ = accuracy score (1 to 5)

C. Decide

1. Choose one main model (best total score).
2. Choose one fallback model (fastest stable model).
3. Save your result in your project notes.

4. Printable Benchmark Template

Copy this section into your project notes and fill it while testing.

Test setup

Model results

Prompt-by-prompt notes

Scoring helper

Use this formula:

$$ ModelScore = 0.35Q + 0.30S + 0.20R + 0.15A $$

Where:

• $Q$ = quality score (1 to 5)
• $S$ = speed score (1 to 5)
• $R$ = reliability score (1 to 5)
• $A$ = accuracy score (1 to 5)

Final decision

5. Recommended Starter Workflow

1. Start with Qwen2.5 4B Instruct for daily use.
2. Switch to Phi-3 Mini for difficult coding/debugging tasks.
3. Use Gemma 2 2B when you need very fast responses.
4. Re-run the benchmark after major LM Studio or model updates.

6. Example of screenshot (03-chat-test.png)

Phi-3 is a family of lightweight 3B (Mini) and 14B (Medium) state-of-the-art open models by Microsoft. The model is licensed under the MIT license.

Last updated: June 2, 2026