Parameters
Also known as · weights · model size
The internal numbers a model learns during training; roughly, its 'knobs'.
Parameters (also called weights) are the adjustable numbers inside a neural network. During training, the model nudges each parameter up or down so its predictions get better. A modern LLM has billions of them, and collectively they encode everything the model 'knows'.
Parameter count is the headline number people use to describe model size — a '70B' model has 70 billion parameters. More parameters generally means more capacity to capture patterns, but it also means more memory, more compute, and higher cost to run. Bigger is not automatically better; data quality and training technique matter just as much.
Parameters are fixed once training ends. When a model 'learns' something new mid-conversation, it isn't changing its parameters — it's just using the information you put in its context window.