# What Makes an AI Model?
Before diving into model names, we need to understand one fundamental concept: Parameters. Simply put, a Large Language Model (LLM) is a massive collection of numbers that have been adjusted during training. By seeing vast amounts of data, the model tunes these parameters to learn patterns in language, code, logic, and information.
Key insight: When you see numbers like
7B, 14B, or 32B next to a model name, they usually refer to the parameter count.
# What Do 7B, 14B, and 32B Mean?
The letter B stands for Billion. So:
7B
≈ 7 Billion Parameters
14B
≈ 14 Billion Parameters
32B
≈ 32 Billion Parameters
70B
≈ 70 Billion Parameters
For example, when you see
Llama-3.1-8B, the 8B means this model has approximately 8 billion parameters.
Common misconception: More parameters doesn't always mean a better model. A 32B model isn't necessarily better than a 14B model for all tasks. Architecture, training data quality, training methods, and post-training tuning all matter significantly.
# What Is A3B?
This is where model names get more interesting. Consider
Qwen3-30B-A3B-Instruct-2507. The 30B-A3B part reveals something crucial about the model's architecture. This is a Mixture of Experts (MoE) model with approximately 30.5 billion total parameters, but only about 3.3 billion parameters are active for each token.
The idea is simple: the model doesn't need to use all its experts for every token. A router selects which experts are most relevant for each input.
1
Question — input arrives
2
Router — decides which experts to use
3
Select Experts — 8 of 128 chosen
4
Process — only ≈ 3.3B params active
5
Answer — output generated
Important:
A3B doesn't mean this model uses the same memory as a 3B model. All 30.5B parameters must still be available for different experts. 30B-A3B ≠ 3B.
# Dense vs. MoE: What's the Difference?
In a Dense model, computations for each token happen across the entire network. In MoE, only selected experts are activated for each token. This allows for large parameter capacity without the full computational cost of activating all parameters.
Dense Model
All parameters active for every token
- Input → All Model Parts → Output
- Full computation per token
- Simpler architecture
MoE Model
Only selected experts active per token
- Input → Router → Selected Experts → Output
- Partial computation per token
- Large capacity, lower compute cost
# What Does Instruct Mean?
One of the most common terms you'll see in model names is Instruct.
Instruct refers to models that have been fine-tuned to follow user instructions. While a base model might be trained primarily for text prediction and continuation, an Instruction-Tuned model is optimized for interactions like:
Summarize
- "Summarize this text"
Code
- "Fix this code"
Translate
- "Translate to Persian"
Create
- "Write a PHP function"
1
Base Model — trained on raw data
2
Instruction Tuning — learns to follow commands
3
Instruct Model — ready for daily use
If your goal is daily use, conversation, or executing commands, the Instruct version is usually the better choice.
# What Does Base Mean?
In contrast to Instruct, you might encounter Base. Base means the base model — the starting point for development, fine-tuning, or specific applications.
Base
Starting point for development
- Text prediction & continuation
- For fine-tuning & research
Instruct
Ready for user interaction
- Follows instructions
- Optimized for conversation
# What Does Coder Mean?
Models with Coder in their name are typically optimized or targeted for programming tasks. Examples include families like Qwen-Coder or DeepSeek-Coder. If your primary focus is programming, a Coder model can be a more logical choice than a general-purpose model of the same size.
PHP
Python
JavaScript
SQL
# What Do Vision or VL Mean?
Some models don't just process text. Terms like Vision, VL (Vision-Language), or VLM refer to models capable of processing visual information. For example, Gemma 4 26B A4B is a new multimodal model from Google that accepts both image and text inputs.
1
Image — visual input
2
Vision Encoder — converts image to tokens
3
Language Model — processes combined input
4
Text Response — describes or analyzes image
Such models can analyze screenshots, charts, or images.
# What Do Reasoning or Thinking Mean?
In newer models, terms like Reasoning, Thinking, or R1 are increasingly common. These models are designed for multi-step reasoning and complex problem-solving.
1
Question — complex problem
2
Analyze — break down the problem
3
Steps — solve step by step
4
Result — verified answer
Note: Reasoning doesn't mean better at everything. For simple questions, a regular model might be faster and more appropriate. But for complex math, programming, or multi-step reasoning, reasoning models can have advantages.
# What Does Distill Mean?
Consider this real model:
DeepSeek-R1-Distill-Qwen-32B. This is a distilled model from the DeepSeek-R1 family. The idea of Knowledge Distillation is that a larger model acts as a teacher, and a smaller model is the student. The goal is for the smaller model to learn some of the larger model's capabilities without needing to be as large.
T
Teacher Model (Large)
DeepSeek-R1 — full reasoning capability
↓
Knowledge Transfer
Behavior / Output / Reasoning patterns
S
Student Model (Small)
Qwen-32B — learns from teacher, much smaller
# What Do Q4, Q5, and Q8 Mean?
Now we reach something crucial for LM Studio, llama.cpp, and local AI users: Quantization. The original model is stored with high precision (e.g., FP32 or BF16). But such a model would be very large. Quantization reduces the precision of weight representation to decrease model size and memory requirements.
32
FP32 — 32-bit float
Highest precision, largest size
16
BF16 / FP16 — 16-bit
Half precision, good balance
8
8-bit
Smaller, minor quality loss
4
4-bit
Much smaller, runs on less RAM/VRAM
Important:
Q4 doesn't mean the model is "4 times worse." It refers to 4-bit quantization level. You're making a trade-off between size and quality.
# What Is Q4_K_M?
Names like
Q4_K_M, Q5_K_M, Q6_K, and Q8_0 are different quantization methods in the llama.cpp ecosystem. Q4_K_M is a specific method from the K-Quant family that can use different bit depths for different parts of the model to create a better balance between size and quality.
Q4_K_M
4-bit, K-Quant Medium
Q5_K_M
5-bit, K-Quant Medium
Q6_K
6-bit, K-Quant
Q8_0
8-bit, near-lossless
# What Is GGUF?
The
.gguf extension is a model file format widely used in the llama.cpp ecosystem and compatible tools. If you've worked with LM Studio, you've likely seen GGUF files many times.
1
Model — trained weights
2
GGUF — converted to file format
3
llama.cpp — loads and runs
4
Local Inference — running on your machine
# What Is Context Length?
Numbers like
8K, 32K, 128K, or 256K usually refer to Context Length — the amount of information the model can maintain in its attention window during a single interaction.
8K
Short context
32K
Medium context
128K
Long context
256K
Very long context
Important:
128K Context means approximately 128 thousand tokens, NOT 128 thousand words. A token might be part of a word, a whole word, a symbol, or a combination of text parts. Token count depends on language and text used.
# What Is a Token?
Language models don't necessarily see text as complete words. Text is first divided into tokens by a tokenizer.
1
Text — raw input string
2
Tokenizer — splits into pieces
3
Tokens — sub-word units
4
Model — processes tokens
# Thinking vs. Non-Thinking
In newer models, even a single model family might have different modes. For example, Qwen3 offers both Thinking and Non-Thinking capabilities.
Non-Thinking
Direct, fast responses
- More direct answers
- Usually faster
- Good for simple tasks
Thinking
Deep reasoning before answering
- More reasoning steps
- Better for complex problems
- Slower but more thorough
# What Is Fine-Tuning?
In fine-tuning, the model is retrained on target data to be better tuned for a specific task. This is completely different from quantization.
1
General Model — broad capabilities
2
Fine-Tuning — retrain on specific data
3
Specialized Model — optimized for a task
Remember: Quantization is about how weights are displayed and stored. Fine-tuning is about changing the model itself through additional training.
# Let's Read a Real Model Name
Let's return to
Qwen3-30B-A3B-Instruct-2507. Now we can read it like a technical ID card:
Q3
Qwen3
Family / Generation of the model
30
30B
≈ 30.5B total parameters
A3
A3B
≈ 3.3B active parameters (MoE)
I
Instruct
Tuned for following instructions
★
2507
Version / Release ID
The official Qwen specifications describe this model with 30.5B total parameters, 3.3B active parameters, 128 experts, and 256K context. A relatively short name contains a lot of information.
# Another Example
Consider
DeepSeek-R1-Distill-Qwen-32B:
R1
DeepSeek-R1
Base model / family
D
Distill
Distilled version — knowledge transferred
★
Qwen-32B
Qwen base model, ~32B size
# A Newer MoE Example
Another real model that clearly shows this naming pattern:
Gemma-4-26B-A4B. Here, 26B refers to the overall model size and A4B to the active parameter count. Google released this model as a multimodal model usable with image and text inputs.
26B
Total Parameters
A4B
Active Parameters (MoE)
Multimodal
Image + Text Input
# Which File Should You Download?
Suppose you see multiple files for a model. Here's a simple guide:
Q4
- Limited memory? Good choice
- Smallest size
Q5
- Good balance
- Size vs. quality
Q6
- Higher quality
- Needs more memory
Q8
- Near-lossless
- Largest size
Important: These aren't absolute rules. Actual quality depends on the model, quantization method, architecture, runtime, and even the type of work you're doing.
# A Common Mistake About A3B and Memory
Suppose you see
30B-A3B and say: "Since only 3B is active, I only need memory for a 3B model."
Wrong
"A3B = only need 3B memory"
- All 30.5B params must be in memory
- Experts are swapped, not removed
Correct
"A3B = less compute per token"
- Full model must be loaded
- But inference is faster
# Quick Reference
Here's a quick guide to the most common terms you'll encounter:
7B / 32B
- Total parameter count
A3B / A4B
- Active params in MoE
MoE
- Mixture of Experts
Instruct
- Tuned for instructions
Base
- Base model
Coder
- Programming focused
Vision / VL
- Image processing
Reasoning
- Multi-step reasoning
Q4 / Q8
- Quantization level
GGUF
- Model file format
128K / 256K
- Context Length
Token
- Text processing unit
# Don't Choose a Model by Name Alone
Perhaps the most important point in this article. When you see
70B, don't immediately say "this is better." When you see Q8, don't say "this is the best." When you see A3B, don't say "this is just a 3B model." And when you see Instruct, don't think "it's the best choice for everything."
A model name is just a clue. For real selection, you need to consider all of these factors together:
Architecture
Parameters
Active Params
Training
Context
Quantization
Hardware
Your Task
# Model Names Are No Longer a Mystery
Next time you visit a model download page and see
Qwen3-30B-A3B-Instruct-2507-Q4_K_M.gguf, you can break it down piece by piece:
Q3
Qwen3
Model family
30
30B
Total parameters
A3
A3B
Active parameters in MoE
I
Instruct
Following instructions
25
2507
Version / Release
Q4
Q4_K_M
Quantization method
★
GGUF
File format for llama.cpp
Suddenly, what looked like a password a few minutes ago becomes a technical ID card for an AI model.
# Final Thoughts
When we work with an AI, we usually only see its output. But behind that short response lie billions of parameters, complex architectures, different experts, tokens, context, quantization methods, and various training stages. Even the model name tells part of this story.
30B tells us about the model's size. A3B tells us about its active portion. Instruct tells us about its training. Q4_K_M tells us about its compression. GGUF tells us about its file format. And when we learn this language, model selection is no longer a random guess.
takeaway.txt
Next time you see a strange model name,
instead of passing it by, read it —
perhaps the entire model story is written right there.
instead of passing it by, read it —
perhaps the entire model story is written right there.
