Parameter, in machine learning, is one of the internal values a model learns during training. Parameters are the adjustable settings that store what the model has learned, and modern models have billions of them.
What it means in plain English
Think of parameters as the millions or billions of tiny dials inside a model. Training is the process of adjusting all these dials so the model produces good outputs. Together, the learned values of all parameters are the model — they encode everything it knows. When you hear a model described as having “70 billion parameters,” that is the number of these internal dials.
More parameters generally allow a model to capture more complex patterns, which is part of why large language models are so capable — and so expensive to train and run.
A simple example
In a simple model predicting a price, a parameter might be the weight given to a house’s size. In a large language model, billions of parameters together encode grammar, facts, and reasoning learned from training.
Why it matters
Parameters are the substance of what a model learns. The parameter count is a rough proxy for a model’s capacity, and understanding parameters clarifies what is actually happening when a model is “trained”: its parameters are being tuned.
Related terms
- Hyperparameter — settings chosen before training, not learned.
- Model — made up of its learned parameters.
- Gradient Descent — the process that adjusts parameters during training.
Frequently asked questions
What is a parameter in a model?
A parameter is a value the model learns during training — such as the weights and biases in a neural network — that determines how inputs are transformed into outputs.
Why do parameter counts matter?
Parameter count is a rough proxy for a model’s capacity; large language models have billions of parameters, which contributes to their capability but also their compute cost.
