Parameters (or 'Weights') are the internal numerical variables a neural network learns during training, dictating how it processes inputs and generates outputs.
Parameters are a model's learned weights: the billions of numbers adjusted during training that encode everything it knows. Counts (7B, 70B, 405B) describe scale: more parameters mean more representational capacity but also more memory and compute per inference. Don't confuse them with hyperparameters (training settings) or sampling parameters (temperature, top-p) you set at inference.
Parameter count is shorthand for the capability/cost trade: it determines GPU memory needs, inference speed, and deployability (an 8B model runs on-device; a 400B model needs serious infrastructure). 2026's efficiency wave, distillation, quantization, better data, keeps shrinking how many parameters a given capability requires.
No: training data quality, architecture, and post-training matter enormously. Modern small models outperform older giants; the right question is capability-per-dollar on your task, measured by evals.
Quantization stores each parameter in fewer bits (8-bit, 4-bit), cutting memory and speeding inference with modest quality cost: the standard trick for fitting big models on small hardware.
Mixture-of-experts models route each token through a subset of expert subnetworks, so a model might hold 400B+ total parameters but compute only tens of billions per token: big-model capability at smaller-model serving cost.