What Are Model Weights?

Weights are the learned numbers inside a neural network that determine how the model behaves.

A model is made of:

• A fixed architecture (layers + math)
• Plus weights (millions or billions of numbers)

The architecture stays the same.
Learning happens by changing the weights.

---

Think of it like this

• Architecture → brain structure
• Weights → synapse strengths
• Training → learning

You don’t change the brain’s shape when you learn — you change the connections.

---

Why weights matter

Same architecture, different weights → different model

That’s why companies like OpenAI don’t release weights:
the weights are the intelligence.

---

Training vs fine-tuning

• Training: start with random weights, learn everything
• Fine-tuning: start with pre-trained weights, adjust them slightly

Fine-tuning = small weight updates

---

Practical rule

If you can download the weights, it’s your model.
If you can’t, you’re using someone else’s intelligence via an API.

---

RAG note

RAG doesn’t change weights.
It adds knowledge outside the model.

---

TL;DR
Weights are the numbers that encode everything an AI model “knows” and “how it thinks”.