Module 3 Lesson 2: Model Naming and Tags
·AI & LLMs

Module 3 Lesson 2: Model Naming and Tags

Decoding the colon. Understanding what 'llama3:8b-instruct-q4_K_M' actually means.

Decoding Model Naming and Tags

When you see a command like ollama run llama3:8b-instruct-q4_K_M, it looks like a cat walked across a keyboard. However, every character in that tag has a specific meaning. Mastering this "code" will help you choose the right model for your specific hardware.

The Structure of a Tag

A model name follows this pattern: [NAME] : [PARAMETER SIZE] - [FLAVOR] - [QUANTIZATION]


1. Parameter Size (The "B")

8b, 70b, 7b, 3b.

  • B stands for Billion.
  • It tells you how many "connections" (parameters) the model has.
  • Rule: More parameters = More Intelligence = More RAM needed.

2. The Flavor (The "What it does")

  • Instruct: Fine-tuned for chat and following instructions. (Use this for 99% of tasks).
  • Text: The raw model. It's good for "predicting the next word" in a book, but bad at answering questions.
  • Vision: Can "see" images. You can upload a photo and ask questions about it.
  • Code: Specialized training for writing Python, JS, C++, etc.

3. Quantization (The Compression)

q4, q5, q8, fp16.

  • This tells you how much the model has been "squashed" to fit on your disk.
  • q4 (4-bit): The industry standard. High speed, small size.
  • q8 (8-bit): Better intelligence, but twice the size of q4.
  • fp16: No compression. Extremely large and slow, but the "truest" version of the model.

The "Latest" Trap

If you run ollama run llama3, you are implicitly running llama3:latest.

Warning: The latest tag is a moving target. If the Ollama team updates the registry, your latest model might change features overnight. For production software, always specify a full tag (e.g., llama3:8b).


Common Examples Decoded

TagWhat it means
llama3:8bThe standard 8 billion parameter Llama 3 model.
llama3:70b-instruct-q4_K_MThe massive 70B model, tuned for instructions, compressed using "K-Medium" 4-bit math.
llava:7b-v1.6-mistral-q4_0A vision-capable model based on Mistral 7B.
phi3:miniThe smallest, fastest version of Microsoft's Phi-3 model.

How to Check Tags on Your Machine

Run ollama list in your terminal. Look at the TAG column. This reflects exactly what version you have locally.


Key Takeaways

  • The Name is the model series.
  • The Tag (after the colon) specifies size, purpose, and compression.
  • Instruct models are what you usually want for chat applications.
  • Parameter count (B) determines the RAM/VRAM requirement.
  • Avoid using latest in production scripts; be specific!

Subscribe to our newsletter

Get the latest posts delivered right to your inbox.

Subscribe on LinkedIn