When an artificial-intelligence company says a model has billions or trillions of parameters, it is describing the number of learned numerical values inside the model.
Google's machine-learning glossary defines parameters as values such as weights and biases that a model learns during training. In a neural network, those values shape how signals are transformed from input to output.
A simple way to think about parameters is as adjustable internal settings learned from data. During training, the model repeatedly changes those settings to reduce error. The final pattern of values helps the model predict, classify or generate outputs.
More parameters can give a model greater capacity to represent complicated patterns, but parameter count is not a direct intelligence score. Two models with similar sizes can perform very differently because of architecture, training data, optimisation methods, context length, tools, post-training and inference techniques.
A very large model can also be expensive. More parameters usually require more memory, computing power and energy to train and run, although techniques such as sparsity, mixture-of-experts designs and quantisation can change the relationship between model size and cost.
Parameters are also different from hyperparameters. Parameters are learned by the model during training. Hyperparameters, such as learning rate, are settings chosen by developers or tuning systems.
This is why a headline saying one model has more parameters than another should be treated as only one technical fact. Real capability should be judged through transparent evaluations, reliability, safety, efficiency and performance on useful tasks.