Parameter

A parameter (also called a weight) is one tiny number inside a model. Models are made of billions or trillions of them, and their job is simple in principle: for each token the model reads, the parameters nudge the model toward its best guess of the next token.

During training, an automated process adjusts each parameter by a hair — billions of tiny nudges over weeks of computing — until the model’s guesses get remarkably good.

Why the number matters:

• ~1 billion parameters → small, fast, can run on a phone or laptop.
• ~7–14 billion → a sweet spot many people run locally.
• 70 billion+ → serious quality, usually cloud-only (or a very serious machine).
• Frontier models → hundreds of billions, trained on massive clusters.

More parameters ≠ always better — but it’s the main lever, and it’s why model names often carry a number (7B, 14B, 70B…).