What does "7B parameters" actually mean?

Every model name you see — Llama 7B, a 70B model, a rumored trillion-parameter giant — is bragging about one number: how many parameters it has. A parameter is simply a single adjustable number stored inside the network, most of them weights that scale the connection between two artificial neurons. "7B" means roughly 7 billion of these knobs. Training is nothing more than nudging all of them, a tiny bit at a time, until the model's outputs stop being wrong.

Think of the whole model as a colossal mixing board. Each parameter is one slider. During training the system plays millions of examples through the board and automatically tweaks every slider to reduce error. Once training stops, the sliders freeze — the finished file you download is literally just that giant list of frozen numbers. When you chat with the model, none of them change; your prompt is only flowing through the fixed settings.

Parameter count is a rough proxy for capacity: more sliders can capture more patterns, which is why bigger models often reason better. But it also sets the cost of running the model. A useful rule of thumb: at common precision, each parameter eats about 2 bytes, so a 7B model needs ~14 GB of memory just to hold its weights — before it processes a single word. That's why the parameter count predicts not only how smart a model feels, but whether it fits on your laptop or demands a data-center GPU.