Parameter

A learned weight in a model, adjusted during training. Parameter count is the headline measure of model size.

A 70B model has 70 billion weights, each stored as a floating-point number. The count determines memory requirements (roughly 2 bytes per parameter at FP16), training cost and - loosely - capability. Larger models are not always better for a given task, and quantisation can reduce the per-parameter cost at the expense of precision.