LM Studio Companion Guide
A companion guide for LM Studio, including a screenshot walkthrough, coding prompts, and mini benchmark checklist.
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.
- Home screen: confirm LM Studio opens without errors.
- Model search: find a model like Qwen2.5 3B Instruct.
- Download panel: choose a quantization, prefer
Q4_K_M. - Model loaded: confirm memory usage looks stable after load.
- Chat tab: send a test prompt and verify response speed.
- Settings tab: set context window to 4096 first.
- GPU settings: enable maximum stable GPU offload.
- API server (optional): enable local OpenAI-compatible endpoint.
- 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.png02-model-search.png03-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
- Pick 3 prompts you will reuse for every model:
- one coding prompt
- one debugging prompt
- one explanation prompt
- 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
- Choose one main model (best total score).
- Choose one fallback model (fastest stable model).
- 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
| Field | Value |
|---|---|
| Date | |
| Machine | |
| LM Studio version | |
| Context window | |
| Temperature | |
| Prompt set used |
Model results
| Model | Quantization | First token (s) | Tokens/sec | Memory pressure | Quality (1-5) | Reliability (1-5) | Accuracy (1-5) | Final score |
|---|---|---|---|---|---|---|---|---|
| Model A | ||||||||
| Model B | ||||||||
| Model C |
Prompt-by-prompt notes
| Prompt type | Model | Short notes | Hallucination level (low/medium/high) |
|---|---|---|---|
| Coding | |||
| Debugging | |||
| Explanation |
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
| Decision item | Selected model | Why |
|---|---|---|
| Main model | ||
| Fallback model | ||
| Re-test date |
5. Recommended Starter Workflow
- Start with Qwen2.5 4B Instruct for daily use.
- Switch to Phi-3 Mini for difficult coding/debugging tasks.
- Use Gemma 2 2B when you need very fast responses.
- 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
Installing Ubuntu 24.04 LTS on Apple Silicon (M1/M2/M3) using UTM
A step-by-step guide for installing Ubuntu 24.04 LTS on Apple Silicon Macs using the UTM virtualization tool, including preparation, VM creation, installation, and post-installation configuration.
LM Studio Local Testing Guide
A guide for testing local LLMs with LM Studio on developer laptops, especially Apple Silicon machines with 8 GB unified memory.