Parameters and Weights

Parameters are the values a machine learning model learns from data during training. In a neural network, most parameters are weights, which set how strongly one input or neuron influences the next, plus biases, which shift each output. When people say a large language model has “billions of parameters”, they are describing how many of these learned numbers it contains.

How it works

Training starts with weights set to random values. The model makes predictions, measures how wrong they are, and nudges every weight slightly in the direction that reduces the error. Repeat this across a large dataset and the weights come to encode patterns: which words tend to follow which, or which features separate spam from genuine email. After training, the weights are fixed and the model uses them for inference.

Parameters are different from hyperparameters, which you choose before training: learning rate, number of clusters, tree depth. They also differ from the “parameters” you pass to an API call, such as language or location, which are just request settings.

How it’s used in SEO and marketing

• Choosing a model. Parameter count is a rough guide to capability, cost and speed. A small embedding model can run on a laptop; the largest LLMs need data-centre hardware.
• Open-weight models. Some providers publish their trained weights, so you can run the model on your own infrastructure or fine-tune it. See open-weight LLMs.
• Fine-tuning. Adjusting a pre-trained model’s weights on your labelled examples, for instance page-type or intent classification.
• Reading model cards to understand what you are paying for and where it can run.

Example

The original BERT paper describes two sizes: BERT-base with 110 million parameters and BERT-large with 340 million. MLforSEO’s meta description guide notes that by 2026 BERT counts as a small model: an encoder built to turn text into numbers, compared with the much larger decoder LLMs that generate text. The same guide also uses hand-set scoring weights (40/45/15) to rank candidate sentences; those are rules you set yourself, unlike learned parameters, so tune them on your pages.

Limitations and things to watch

• More parameters do not guarantee better results on your task. Training data, architecture and fine-tuning matter as much.
• Providers often do not disclose parameter counts for proprietary models, so comparisons rely on benchmarks.
• Weights encode whatever patterns the training data contained, including biases.
• Weights cannot be read like rules. Explaining why a model made a decision needs separate explainability methods.

Learn it on MLforSEO

• How to Automatically Generate Meta Descriptions with Python: From BERT to LLMs
• The Characteristics of a Machine Learning Problem
• Introduction to Machine Learning for SEO course, including pre-trained vs fine-tuned vs self-trained models.