Weights

Weights — also called parameters — are the internal numbers that define how a neural network behaves. During training the model adjusts them across billions of examples until it gets good at its task; "7 billion parameters" or "70 billion weights" describes how many of these numbers a model has. The model doesn't store facts the way a database does — it encodes patterns in the weights.

Size carries trade-offs: more weights mean more capacity for complex patterns, but more compute to train and run — a 7B model can fit on a laptop, while a 70B model needs serious hardware. Fine-tuning works by updating a subset of these weights rather than starting from scratch.

Weights are what you download in a model file: the trained knowledge, separate from the architecture that defines how they connect. Hallucinations and errors usually trace back to a pattern in the weights that doesn't quite fit the situation.