Module 3: LM Studio - Introduction

Downloading Models (Model Zoo)

Capsule overview

LM Studio without a model is like a browser without internet. In this capsule you'll download your first local LLM model.

You'll learn to:

  • Explore the Model Zoo (model catalog)
  • Understand quantization (Q4, Q8, etc.)
  • Download Mistral 7B (recommended first model)
  • Manage models (delete, switch)

Time: 20 min + download (10-30 min depending on your internet)
Difficulty: Low


🎯 Objectives

  • ✅ Explore the Model Zoo
  • ✅ Understand quantization levels
  • ✅ Download 1+ model (Mistral 7B)
  • ✅ Verify the model works

🔍 Step 1: Explore the Model Zoo

1.1 Open the Discover tab:

In LM Studio, click the 🔍 Discover tab (top).

You'll see a catalog of available models with information about size, license, and a download button.


1.2 Available filters:

By size:

  • Small (< 5GB) → 7B parameters
  • Medium (5-15GB) → 13B parameters
  • Large (> 15GB) → 70B+ parameters

By quantization:

  • Q4 (4-bit) → Faster, lower quality
  • Q5 (5-bit) → Balance
  • Q8 (8-bit) → Slower, better quality

By task:

  • Chat/Instruct → Conversational
  • Code → Programming
  • Embedding → Vectors (RAG)

📊 Step 2: Understand Quantization

What is quantization?

Reducing the precision of the model's weights to use less memory.

In LLMs:

  • Original model (FP16): 100% quality, 14GB RAM
  • Q8 (8-bit): 95% quality, 7GB RAM
  • Q4 (4-bit): 85-90% quality, 4GB RAM

Quantization levels:

QuantizationSizeQualitySpeedRAM (7B model)
Q87GB95%Medium10GB
Q55GB90%Fast8GB
Q44GB85-90%Very fast6GB

Recommendation to start: Q4 (perfect balance)


📥 Step 3: Download Mistral 7B

3.1 Search for the model:

In the Discover tab, search for: Mistral 7B Instruct

3.2 Choose the quantization:

Choose Q4_K_M (recommended):

  • Size: ~4.37 GB
  • Quality: 85-90%
  • Speed: Fast

Click [Download]

Time: 10-30 minutes (depending on your internet)


📁 Step 4: Model Management

4.1 View downloaded models:

Home tab → My Models

You'll see a list of downloaded models with options to Load and Delete.

4.2 Delete a model:

If you need to free up space, click [Delete] and confirm.


🧪 Step 5: First Load and Test

5.1 Load the model:

  1. Home → My Models
  2. Click [Load] on Mistral 7B
  3. Wait 5-30 seconds

Indicator: 🟢 in the top bar when it loads correctly


5.2 Test in Chat:

  1. Click the 💬 Chat tab
  2. Type: Hi, what model are you?
  3. Enter

Expected response:

I'm Mistral 7B, a language model developed by Mistral AI.
I'm designed to answer questions and hold conversations.

If you see a response: ✅ The model works correctly!


📚 Recommended Models

To start (7B):

1. Mistral 7B Instruct v0.2

  • Size: 4.37 GB (Q4)
  • Best for: General chat, Spanish, English
  • License: Apache 2.0 ✅

2. Llama 2 7B Chat

  • Size: 3.83 GB (Q4)
  • Best for: General chat
  • License: Llama 2 Community

If you have 16GB+ RAM (13B):

3. Llama 2 13B Chat

  • Size: 7.33 GB (Q4)
  • Best for: Better quality than 7B
  • Trade-off: Slower

For code (7B):

4. Code Llama 7B

  • Size: 3.8 GB (Q4)
  • Best for: Python, JavaScript, etc.
  • Specialized in programming

⚙️ Advanced Models

If you have 32GB+ RAM (70B):

Llama 2 70B Chat

  • Size: 40 GB (Q4)
  • Quality comparable to GPT-3.5
  • Very slow without a powerful GPU

Only recommended if:

  • You have 32GB+ RAM
  • A GPU with 16GB+ VRAM (NVIDIA 4090, A100)
  • Or patience for 30-60s per response

🔄 Switching Models

Change the active model:

  1. Home → My Models
  2. Click [Unload] on the current model
  3. Click [Load] on the new model
  4. Wait for it to load

Tip: You can only load 1 model at a time (RAM limitation)


📊 Model Comparison

ModelSize (Q4)QualitySpeedSpanishCode
Mistral 7B4.4 GB⭐⭐⭐⭐⚡⚡⚡✅ Excellent⭐⭐⭐
Llama 2 7B3.8 GB⭐⭐⭐⚡⚡⚡⭐⭐⭐ Good⭐⭐⭐
Llama 2 13B7.3 GB⭐⭐⭐⭐⚡⚡⭐⭐⭐⭐⭐⭐⭐⭐
Code Llama 7B3.8 GB⭐⭐⭐⚡⚡⚡⭐⭐⭐⭐⭐⭐⭐

🐛 Troubleshooting

Download very slow

Cause: Server overload or slow internet

Solution:

  • Wait (it downloads in the background)
  • Or download from Hugging Face directly and place it in the models folder

Error "Not enough RAM"

Cause: The model is too large for your hardware

Solution:

  • Use a lower quantization (Q4 instead of Q8)
  • Close other apps
  • Download a smaller model (7B instead of 13B)

Model doesn't appear after downloading

Cause: Corrupt download

Solution:

  1. Delete the model
  2. Re-download
  3. Check the file size on disk

📊 Summary

What you did:

  1. Explored the Model Zoo:

    • A catalog of 100+ models
    • Filters by size, quantization, task
  2. Understood quantization:

    • Q4: Fast, 4GB (recommended)
    • Q8: Slow, 8GB (better quality)
  3. Downloaded Mistral 7B:

    • 4.37 GB (Q4)
    • Apache 2.0 license
    • Spanish + English
  4. Tested the model:

    • Loaded into RAM
    • First conversation
    • Verified it works

Checklist:

  • You downloaded 1+ model (Mistral 7B recommended)
  • The model appears in Home → My Models
  • You loaded the model successfully (🟢)
  • You tested it in the Chat tab
  • You got a coherent response

If all ✅: Ready to use the full interface!


🔗 Additional resources

  1. Hugging Face Model Hub - All GGUF models
  2. Mistral AI - Official Mistral docs
  3. Quantization Guide - Technical details

➡️ Next step

Next capsule: 04-the-chat-interface.md

Now that you have a model downloaded, you'll:

  • Explore the Chat interface
  • Adjust parameters (temperature, etc.)
  • Use the playground
  • Compare different prompts

Time: 20 minutes
Difficulty: Low


Estimated time: 20 min + download (10-30 min)
Next: 04-the-chat-interface.md