Glossary — Language models

Parameter

A parameter is one number inside a model, adjusted during training — large models have hundreds of billions of them.

1 tokensParameter

Simply put#

When you hear "a 70 billion parameter model", that is 70 billion numbers settled into place by training. Those numbers are everything the model "knows".

Example#

More parameters generally means a more capable model, but also a more expensive and slower one. Smaller models can beat larger ones on a narrow task.

Why it matters to you#

Parameter count is a popular marketing number, but on its own it says nothing about quality. The quality of the training data matters as much or more.

Related terms