Large language models concept

Parameters versus weights

Reference

What are model weights? They are learned numerical values in a neural network that help shape how it processes input. Parameters is the broader term for learned values in a model, including weights and, in common neural networks, biases. So weights are one kind of parameter in the precise sense. In casual discussion, people sometimes use “weights” to mean all the model’s learned parameters.

Reviewed 2026-10-09

The two things

In a neural network, a parameter is a learned value that helps determine the model’s behavior. Weights are values that multiply signals as they pass through the network; biases are learned values added to those signals. Both can be parameters, so weights are part of the parameter set.

A simple linear model makes the distinction visible: a feature’s weight controls its contribution to the prediction, while the bias shifts the prediction. Both values are calculated during training.

The key difference

The precise distinction is that parameter names the broader group of learned values, while weight names one kind of value in that group. In neural networks, the group can include weights and biases. A model’s parameter count therefore counts learned values, not necessarily only weight values.

Usage can be looser: people sometimes call all the learned parameters “weights,” with the bias understood. That shorthand is common enough that the surrounding explanation matters. Also, a hyperparameter is different: it is a setting chosen or tuned for training, such as the learning rate, rather than a value the model learns.

When each term is useful

Use parameters when you mean the full set of learned values or are describing how many learned values a model has. Use weights when you mean the multiplicative values applied to signals.

For example, Google writes a linear prediction as y′ = b + w₁x₁: w₁ is the feature’s weight, while b is the bias. Both are parameters calculated during training. If a source uses “weights” for the full learned set, check whether it is using the word as shorthand for weights and biases together.

Common confusions

“Model parameters” does not mean every setting associated with training or using a model. Weights and biases are learned parameters; choices such as learning rate and batch size are hyperparameters that a designer chooses or tunes.

A weight and a bias are not interchangeable: a weight multiplies a signal, while a bias adds an offset. But informal explanations may call both groups “weights.” When a precise count or description matters, check whether the writer means only weights or all learned parameters.

Questions people ask

What are LLM parameters?

They are learned values that help determine a language model’s output. In neural networks, these include weights and biases; training hyperparameters such as learning rate are a separate category.

Are parameters and weights the same thing?

Precisely, weights are one kind of parameter; biases can also be parameters. In casual usage, “weights” sometimes means the whole set of learned parameters.

Does a model’s parameter count count only weights?

Not necessarily. The count can include biases as well as weights, so check how the source is using “parameters” and “weights.”

Is the learning rate a model parameter?

It is usually called a hyperparameter: a training setting chosen or tuned by the designer. Weights and biases are learned during training.

Learn more

Linear regression 80 min
A worked equation shows a feature weight and bias as separate model parameters.
Speech and Language Processing, Chapter 6
The neural-network summary identifies weights and biases as parameters learned by gradient descent.
What Are LLM Parameters?
An LLM-focused explanation distinguishes learned weights and biases from hyperparameters.

What changed

Sources

Part of Large language models: How a model learns and continues patterns