This is week 22 of my “I Used It for a Week” series. Last week I reviewed Codex CLI, OpenAI’s terminal agent. This week: the tool that runs AI models on your own hardware.

LM Studio is not a coding agent. It’s a model runner. You download models, run them locally, and use them through an API. But I’ve been using it as the backend for Aider and MiMo Code, and that workflow changed how I think about AI coding.

After a week of running models locally, I think LM Studio is the best entry point for anyone who wants to try local AI without the hassle.

How It Works

1. Download LM Studio from lmstudio.ai
2. Browse the model library
3. Download a model (GGUF format)
4. Start the local server
5. Point your tools at http://localhost:1234

The whole process takes about 10 minutes, including model download time.

I tested with several models:

• Llama 4 Maverick (400B, 17B active)
• Qwen 3.6 27B
• DeepSeek V4 Flash
• Ling 2.6 Flash

Day 1: First Impressions

I downloaded Llama 4 Maverick (Q4 quantization, ~60GB) and started the server. It loaded in about 2 minutes on my M2 Pro with 32GB unified memory.

The speed was impressive: 25-35 tokens per second. Fast enough for interactive use.

I connected Aider to it and started coding. The experience was surprisingly good. The model understood my codebase, generated reasonable code, and ran tests.

The key difference from cloud APIs: zero latency on startup, zero cost per token, and complete data privacy.

Day 2-3: Model Selection

LM Studio supports hundreds of models. The challenge is picking the right one for your hardware.

For 16GB RAM:

• Qwen 3.6 7B (Q4): Fast, good for simple tasks
• Llama 4 Scout (Q4): Decent quality, larger context

For 32GB RAM:

• Qwen 3.6 27B (Q4): Best balance of quality and speed
• Llama 4 Maverick (Q4): More capable, slower

For 64GB+ RAM:

• DeepSeek V4 Flash (Q4): Excellent coding performance
• Ling 2.6 Flash (Q4): Good balance of quality and speed

The sweet spot for most developers is Qwen 3.6 27B on 32GB RAM. Good quality, fast enough for interactive use, and doesn’t break the bank.

Day 4-5: The API Server

LM Studio’s killer feature is the built-in API server. It’s OpenAI-compatible, meaning any tool that works with OpenAI’s API works with LM Studio.

I connected:

• Aider: aider --model lmstudio/qwen3.6-27b
• MiMo Code: Configure custom backend
• Continue.dev: Add LM Studio as a provider

The experience was seamless. The tools work the same as with cloud APIs, just with local inference.

The cost: $0. Zero tokens per second. Zero monthly subscription. Just electricity.

What Blew Me Away

Zero cost

Running models locally costs nothing beyond electricity. For a developer who codes 8 hours a day, that’s significant savings compared to $50-100/month on API costs.

Data privacy

Your code never leaves your machine. No cloud APIs, no data transmission, no security concerns. For proprietary code, this matters.

Speed

Modern quantized models run fast enough for interactive use. 25-35 tok/s on consumer hardware is more than adequate.

Model variety

LM Studio supports hundreds of models. You can switch between Llama, Qwen, DeepSeek, Mistral, and others without changing your workflow.

What Frustrated Me

Quality ceiling

Local models are good but not frontier-level. Qwen 3.6 27B is comparable to GPT-4o on many tasks, but it’s not Claude Sonnet 5 or GPT-5.6 Sol quality. For complex tasks, cloud APIs still win.

RAM requirements

Running large models requires significant RAM. Llama 4 Maverick needs 64GB for comfortable use. Most developers don’t have that much.

No coding-specific features

LM Studio is a general-purpose model runner. It doesn’t have the coding-specific optimizations of Claude Code, Aider, or MiMo Code. You need to pair it with a coding tool.

Startup time

Loading a large model takes 1-3 minutes. If you switch models frequently, this adds up.

LM Studio vs Cloud APIs

• Cloud APIs (Claude, GPT, Gemini): Best quality, no hardware requirements, but ongoing costs and data leaves your machine.
• LM Studio: Zero cost, complete privacy, but lower quality and requires decent hardware.

They’re complementary. Use LM Studio for daily coding where quality is adequate. Use cloud APIs for complex tasks where you need frontier quality.

Would I Keep Using?

Yes. LM Studio is now part of my daily workflow. I run Qwen 3.6 27B for routine coding tasks and switch to cloud APIs only when I need frontier quality.

Rating: 9/10 — The best entry point for local AI. Zero cost, great model selection, and seamless API compatibility.

FAQ

What is LM Studio?

LM Studio is a desktop application that lets you download and run AI models locally. It provides an OpenAI-compatible API server, making it easy to use with coding tools like Aider and Continue.dev.

What hardware do I need?

Minimum: 16GB RAM for 7-14B models. Recommended: 32GB RAM for 27B models. 64GB+ for 400B+ models. Apple Silicon Macs work well.

Can I use LM Studio with Aider?

Yes. LM Studio provides an OpenAI-compatible API at http://localhost:1234. Point Aider at it with --model lmstudio/model-name.

What models work best?

For most developers, Qwen 3.6 27B on 32GB RAM is the sweet spot. It offers good quality and fast inference. For simpler tasks, 7-14B models work well on 16GB.

Is LM Studio free?

Yes. LM Studio is free for personal use. Commercial use requires a license.

Related: LM Studio Complete Guide · How Much VRAM for AI · Best Local AI Models 2026

Next week: I Used Langfuse for a Week — open-source LLM observability.