Weights are the adjustable values inside a neural network that determine how strongly one neuron influences another. Learning, for a neural network, is the process of tuning these weights.
What it means in plain English
Every connection between neurons in a network has a weight — a number that controls how much signal passes along it. A large weight means a strong influence; a small one, weak. During training, these weights are gradually adjusted so the network produces better outputs. Together, all the weights encode everything the model has learned; they are, in effect, the model’s knowledge. “Weights” and “parameters” are often used interchangeably.
A model with billions of parameters has billions of these weights.
A simple example
When people talk about downloading a model’s “weights,” they mean the file containing all these learned values — the trained substance of the model that lets it make predictions.
Why it matters
Weights are what a neural network actually learns and stores. Understanding them clarifies what “training” really adjusts, what an “open weights” model shares, and why the values inside a model are its most valuable asset.
Related terms
- Parameter — essentially another word for weights.
- Neural Network — where weights live.
- Gradient Descent — the process that adjusts weights.
Frequently asked questions
What are weights in a neural network?
Weights are the learnable values on the connections between neurons that determine how strongly one neuron’s output influences the next — they encode what the model has learned.
How are weights set?
They start random and are adjusted during training (via backpropagation and gradient descent) to reduce error, ending up as the trained model’s “knowledge.”
