LM Studio Local Testing Guide

LM Studio Local Testing Guide

This guide helps you run and test a local LLM with LM Studio on a developer laptop, especially Apple Silicon machines with 8 GB unified memory.

1. What LM Studio is for

LM Studio lets you:

• Download and run LLMs locally (for example Qwen, Phi, Gemma, Llama)
• Chat with models using a desktop UI
• Expose a local OpenAI-compatible API endpoint for your apps
• Keep prompts and responses on your machine (offline usage is possible after model download)

2. Hardware expectations (important)

For an 8 GB Mac, stay in the 2B to 4B model range for day-to-day work.

• 2B to 4B models: usually fast and stable
• 7B+ models: can work, but often need heavy quantization, smaller context, and slower generation

3. Recommended models (June 2026)

If your main goal is coding, start with:

1. Phi-3 Mini 3.8B Instruct
2. Qwen2.5 4B Instruct

4. Install and first run

1. Download LM Studio for your OS from the official site.
2. Install and open LM Studio.
3. Go to model search and download one model from the list above.
4. Prefer Q4_K_M quantization for 8 GB systems.
5. Start a chat and run a quick sanity test prompt.

Example test prompt:

Explain what this TypeScript function does and suggest one improvement:
function sum(a: number, b: number) { return a + b; }

5. Good default settings for 8 GB Macs

• Quantization: Q4_K_M (fallback: Q4_0 if memory is tight)
• Context window: 4096 to 8192 tokens
• GPU offload: as high as LM Studio allows without instability
• Flash attention: enable if available

Avoid on 8 GB as a daily setup:

• Most 7B and 8B models with large context windows
• 14B+ models
• High-bit quantizations unless model is very small

6. Suggested model lineup for web/dev work

For Python, TypeScript, React/Next.js, SQL/Prisma:

1. Qwen2.5 4B Instruct Q4_K_M (main assistant)
2. Phi-3 Mini 3.8B Instruct Q4_K_M (coding specialist)
3. Gemma 2 2B Instruct Q4_K_M (fast fallback)

7. Troubleshooting quick wins

• If generation is slow: reduce context window first, then switch to a smaller model.
• If LM Studio becomes unstable: lower model size or use a lower-memory quantization.
• If answers are low quality: try a better 4B model before increasing context.
• If you need fully offline use: download models first, then disconnect from network.

8. Optional: MLX on Apple Silicon

If you want maximum Apple Silicon performance, MLX-based model runtimes can be faster and more memory-efficient than GGUF in many cases. LM Studio is usually the easiest starting point, while MLX is a stronger performance path for advanced users.

Last updated: June 1, 2026

Continue with the companion page: LM Studio Companion: Walkthrough, Prompts, and Benchmarks