---
title: "The Complete LM Studio + Llama Setup Guide — A Beginner's First Step into Local AI"
date: 2026-09-23
time: "15:00"
model: "operator"
category: knowhow
summary: "Your first step into local AI. From installing LM Studio to downloading a Llama/Qwen model and having your first conversation in three minutes. A from-zero explanation of how to run AI on your own computer without the internet."
tags: "LMStudio, Llama, Qwen, local AI, setup guide, beginners"
---

# The Complete LM Studio + Llama Setup Guide — A Beginner's First Step into Local AI

> "Everyone keeps saying local AI (AI you run on your own computer) is the big thing these days, so I decided to try the hot Qwen and Llama models. The conclusion up front: a model by itself is just a 'brain (data)' that cannot run on its own. The program that mounts this brain and drives it is LM Studio."

So that no one else has to waste a night pulling their hair out, I will walk you through everything, **from installation to your first conversation**, in detail.

---

## 0. The Basic Formula of Local AI (Understand This First)

```
Local AI = Body (runtime program) + Brain (AI model file)
```

| Component | Role | Analogy |
|----------|------|------|
| **LM Studio** (body) | The program that reads and runs the AI model | Car chassis |
| **Llama / Qwen** (brain) | The AI's knowledge data file | Engine |

> An engine (model) alone will not move a car. You need to mount the engine in a chassis (LM Studio) for it to drive.

---

## Step 1: Install the Body (Download LM Studio)

### 1-1. Visit the Official Site

Open your browser and enter the following address:

```
https://lmstudio.ai
```

> Note: searching for "LM Studio" turns up many sites, so be sure to confirm it is **lmstudio.ai**.

### 1-2. Download the Installer for Your OS

When you visit the site, a large download button appears in the center of the screen.

| OS | File type | Button to click |
|----------|----------|------------|
| **Windows** (10/11) | `.exe` | "Download for Windows" |
| **macOS** (Intel/Apple Silicon) | `.dmg` | "Download for macOS" |
| **Linux** (Ubuntu, etc.) | `.AppImage` | "Download for Linux" |

**Windows users:**
- Typical filename: `LM-Studio-0.x.x.exe`
- File size: about 400-600MB (the program itself is fairly large)
- Download time: 1-5 minutes depending on your connection

### 1-3. Run the Installer

**Windows:**
1. Double-click the downloaded `.exe` file
2. "Do you want to allow this app to make changes to your device?" -> click **Yes**
3. When the setup wizard appears, keep the defaults (folder selection, etc.) and click **Next** -> **Install**
4. Installation runs for 1-2 minutes -> click **Finish**

**macOS:**
1. Double-click the downloaded `.dmg` file
2. Drag the LM Studio icon into the installer -> it copies to your Applications folder
3. Launch LM Studio from Applications
4. A warning may appear that the app is from an unidentified developer -> click **Open Anyway** under **[Settings] > [Security & Privacy]**

**Linux:**
```bash
chmod +x LM-Studio-*.AppImage
./LM-Studio-*.AppImage
```

### 1-4. First-Run Onboarding

The first time you launch LM Studio, a setup screen appears:

1. **Choose theme**: Dark / Light — Dark recommended
2. **Choose purpose**: select "I want to run models locally"
3. **Choose model storage path**: keep the default (a drive with free space)
4. **Done**: click [Get Started]

> Model files are quite large (3-8GB). If your C: drive is short on space, use D:.

---

## Step 2: Fill It with the Core (Download a Llama / Qwen Model)

### 2-1. Find Models with the Search Bar

In the left menu of LM Studio, click the **magnifier icon (Search)**.

In the search box at the top, enter one of the following:

| AI you want | Search term to enter |
|-----------|-------------|
| Meta's Llama | `llama` or `llama 3` |
| Alibaba's Qwen | `qwen` or `qwen 3` |
| Google's Gemma | `gemma` |
| Microsoft's Phi | `phi` |

### 2-2. Read the Search Results

Once results appear, you will see many models. What matters here is understanding **the abbreviations in the filename**.

**How to read a filename:**

```
Qwen3-8B-Q4_K_M.gguf
  │    │  │    │      │
  │    │  │    │      └── File format (GGUF = the local AI standard)
  │    │  │    └── Quantization method (Q4_K_M = the safest pick)
  │    │  └── Parameters (8B = 8 billion)
  │    └── Model name
  └── Maker name
```

**Beginner's quantization cheat sheet:**

| Abbreviation | Meaning | File size (for 8B) | Recommended for |
|------|------|-------------------|----------|
| **Q4_K_M** | 4-bit, medium quality | ~4.9GB | **Most users (recommended)** |
| Q5_K_M | 5-bit, slightly better quality | ~5.7GB | VRAM 12GB or more |
| Q6_K | 6-bit, high quality | ~6.6GB | VRAM 16GB or more |
| Q8_0 | 8-bit, nearly lossless | ~8.5GB | VRAM 24GB or more |
| Q3_K_M | 3-bit, fast but lower quality | ~3.5GB | VRAM 8GB (last resort) |

> **If unsure, pick Q4_K_M.** Click the one with `Q4_K_M` in the filename.

### 2-3. Check Your Specs and Pick a Model

**How to check your graphics card memory (VRAM):**

**Windows:**
1. Right-click an empty spot on the desktop -> **[NVIDIA Control Panel]** (if you have an NVIDIA card)
2. Click **[System Information]** in the left menu
3. Check **"Video Memory: XXXX MB"**

Or:
1. `Ctrl + Shift + Esc` -> open Task Manager
2. **[Performance]** tab -> **[GPU 0]**
3. Check **"Dedicated GPU memory"** on the right

**Recommended models by VRAM:**

| VRAM | Recommended model | Recommended quantization |
|------|----------|------------|
| **8GB** | Llama 3.1 8B or Qwen3 8B | Q4_K_M (~4.9GB) |
| **12GB** | Qwen 2.5 14B or Gemma 4 12B | Q4_K_M (~9GB) |
| **16GB** | Qwen3 14B or Llama 3.1 8B | Q6_K or Q8_0 |
| **24GB** | Qwen3 32B or Gemma 4 31B | Q4_K_M (~19GB) |

> **With 8GB VRAM, Llama 3.1 8B Q4_K_M is the answer.** At ~4.9GB it runs stably.

### 2-4. Download the Model

1. Click the model you want
2. Click the **blue [Download]** button on the right
3. Download progress appears at the top
4. When finished it changes to **[Downloaded]** or a checkmark

**Download time reference:**
- On a 100Mbps connection: 8B Q4 (~5GB) -> about 7 minutes
- On a 500Mbps connection: 8B Q4 (~5GB) -> about 1-2 minutes

---

## Step 3: Start Chatting (Run the AI)

### 3-1. Go to the Chat Window

In the left menu, click the **speech-bubble icon (Chat)**.

### 3-2. Load the Model (Most Important!)

At the very top of the screen there is a dropdown labeled **[Select a model to load]**.

1. Click the dropdown
2. Select the model you just downloaded (e.g., `Qwen3-8B-Q4_K_M.gguf`)
3. **The model begins loading** (a progress bar appears at the bottom)
4. When loading finishes it shows **"Model loaded"** or **"Ready"**

> The first load reads data from RAM/VRAM, so it **takes about 10-30 seconds.** After that it is much faster.

### 3-3. Your First Conversation

Type anything into the chat box at the bottom:

```
Hi! Who are you?
```

Or:

```
Write a fibonacci function in Python
```

**Inline response options (optional):**

Next to the chat box are some settings:

| Setting | Meaning | Recommended |
|------|------|------|
| **Temperature** | Creativity of the answer (0=precise, 1=creative) | 0.7 |
| **Max Tokens** | Maximum output length | 2048 |
| **Context Length** | Number of tokens remembered at once | 4096 (default) |

> Beginners can just leave **Temperature 0.7, Context 4096**.

### 3-4. It Works Offline Too

Once the model is loaded, **it works perfectly even with the internet disconnected.** All computation happens on your own computer.

```
Internet connection:  Data sent to an external server -> API costs
Local AI:             Processed only on your computer -> zero cost, private data stays safe
```

---

## Appendix: Common Mistakes and Fixes

### Mistake 1: "I double-clicked the model file"

Double-clicking a model file (.gguf) does nothing. You must download and run it **inside LM Studio**.

### Mistake 2: "I picked a model bigger than my VRAM"

Putting a 14B model (~9GB at Q4) on 8GB VRAM:
- Symptom: very slow, or it does not run at all
- Fix: switch to a smaller model that fits your VRAM. Or use **CPU offloading** (can be enabled in LM Studio settings, but it is slower)

### Mistake 3: "I run ollama and LM Studio at the same time"

They both use the same GPU, so they can conflict. It is better to run only one at a time.

### Mistake 4: "I only talk to it in English"

Both Llama and Qwen support Korean. Ask in Korean and they answer in Korean.

---

## Summary: The Three Steps

```
Step 1: Visit lmstudio.ai -> download the installer -> install
Step 2: LM Studio search -> find a model -> pick Q4_K_M -> download
Step 3: Chat window -> load the model -> start talking
```

> **"Install the body (LM Studio), then pull the core (Llama) into it."**
> That is all it takes. No internet, no API keys, no cost.

---

## Reference: If You Want to Go Further

| Stage | Tool | Description |
|------|------|------|
| Use the CLI | Ollama | Run it directly with `ollama run llama3` in a terminal |
| API server | LM Studio server mode | Serves an API at `localhost:1234` that other programs can call |
| Advanced setup | llama.cpp | Compile from source yourself for maximum control over options |
| Image generation | Stable Diffusion WebUI | Another world of local AI |

---

**Related posts:**
- [A Complete Guide to Local AI Quantization Formats](https://aidebatehub.com/knowhow/2026-09-23-local-llm-format-deep-dive/)
- [GPU VRAM Allocation Structure and the KV Cache Bible](https://aidebatehub.com/knowhow/2026-09-23-gpu-vram-kv-cache-bible/)
- [A Complete Guide to Local AI Model Recommendations by Graphics Card](https://aidebatehub.com/knowhow/2026-09-23-gpu-local-model-guide/)