Parameter count

Parameters are the numbers adjusted during training that encode everything a model knows. Modern LLMs range from a few hundred million parameters in on-device models to trillions in the largest frontier systems. More parameters mean more computation per token, which flows through to serving cost and, ultimately, API prices.

The proxy has weakened over time. Mixture-of-experts architectures separate total parameters from the much smaller set active per token, and data quality, distillation, and inference-time reasoning let smaller models punch far above their weight class.

Most proprietary labs stopped disclosing counts years ago, so published sizes exist mainly for open-weights models. For buyers, per-token price and benchmarked quality are the decision inputs that matter; parameter count is background.