A parameter is an internal variable a model learns during training and uses when producing predictions. In neural networks these are the weights and biases. They start out randomly assigned, get nudged at every step by the error signal, and in the end everything the model knows lives in those numbers.
It should not be confused with a hyperparameter. A parameter is what the model learns for itself, a hyperparameter is a setting chosen up front by whoever runs the training: learning rate, layer count, batch size. One comes from the data, the other from a decision.
With language models the parameter count doubles as a measure of model size. The difference between a seven billion parameter model and a hundred billion one shows up in capability and in running cost. Parameter count translates directly into memory requirements, so it also decides which models can run on local hardware.
A concrete case: a team is choosing a model for customer support. The large model answers better but every call is expensive and slow. Fine-tuning a seven billion parameter model on their own data gets close to the large model on that narrow task at noticeably lower cost.
Parameter count alone is not a quality signal. Two models of the same size can perform very differently depending on the quality of their training data.
