Parameters are the millions or billions of learned numbers that make up a model's knowledge.
DefinitionWhat it means
Parameters are the numerical weights inside a neural network that are adjusted during training and that collectively encode everything the model has learned. Modern frontier models range from a few billion to well over a trillion parameters, and parameter count is one of the main levers, alongside data and compute, that scaling laws use to predict capability.
Why it mattersWhy you should care
Parameter count is a rough, imperfect proxy for capability and cost: more parameters generally means better performance but also higher inference cost, more memory, and higher latency, which is why teams choose between a huge flagship model and a smaller, cheaper one based on the task's actual difficulty. Vendors increasingly avoid publishing exact parameter counts for competitive reasons, so practitioners often reason about model tiers instead of raw numbers.
At a glanceSee it
Not every number in a model is learned — parameters are fit to the data, while hyperparameters like learning rate are knobs humans choose beforehand.
Each parameter value is discovered by a repeating loop — predict, measure error, assign blame, then nudge — which can overshoot into memorizing noise instead of the real signal.
Where you see itIn the wild
- Model cards stating parameter count, such as 70B or 405B
- Cost comparisons between small and large parameter tiers
- Discussions of quantization to shrink parameter memory footprint