Parameters means the numbers inside an AI model that it sets for itself while it learns. Together they hold everything the model picked up in training. Weights are the same idea: the main kind of parameter, and often used as another word for the whole set.
In everyday English, “parameters” means limits or rules, as in “within the parameters of the job”. In AI it means the learned numbers.
Think of a huge mixing desk with millions of sliders. Training is someone nudging every slider, a little at a time, until the sound comes out right. Nobody sets each slider by hand, and no single slider means much on its own. The final position of all the sliders is the model. The sliders are the parameters.
In a sentence: “The smaller model has far fewer parameters, so it fits on a phone, but it knows less than the big one.”
What the word means
The plain point: parameters are learned, not written by a programmer.
Google’s machine learning glossary defines parameters as “The weights and biases that a model learns during training.”1
A US government report on open models says the values of the weights, “and therefore the behavior of the model,” are set by training on many examples.2
The examples it learns from are its training data.
A simple example
The plain point: the count tells you how big a model is.
Models are often described by how many parameters they have, written with a B for billions. In a 2022 study, researchers compared a model with “70B parameters” against a larger one with 280B.3
Size has a practical side. A model’s file has to fit in your computer’s memory to run at home, so bigger models need bigger machines. Run AI on your own computer shows what fits.
Why it matters to you
The plain point: more parameters does not simply mean a better model.
In that same 2022 study, the smaller model, trained on more data, did better.4 Those models are old now. The lasting point: data and training methods count as well as size. Compare models on your own tasks.
Weights also matter for who controls a model. When a company publishes them, anyone can download and run the model. That is open weights.
A parameter count is a size label, like engine size on a car. It hints at what a model needs to run, not how well it will do your job.
Often confused with
- Settings you choose: options like temperature are also called parameters in AI tools, but you set them and the model does not learn them. Google’s glossary calls the training-time kind hyperparameters.5 Gemini temperature deprecated shows one such setting changing.
- Tokens: the pieces of text a model reads and writes. Parameters are inside the model; tokens pass through it. See token meaning.
- Bias: here it is just a kind of parameter, a number. It is not the same as unfair slant in answers, covered in AI bias meaning.
Questions people ask
What does “70B” mean in a model name?
It means about 70 billion parameters. The Chinchilla research paper, for example, describes a model trained “with 70B parameters”.3
Are more parameters always better?
No. How much and how well a model is trained matters too. In the same paper, the 70-billion-parameter Chinchilla beat much bigger models: it “uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B)”.4
Parameters vs hyperparameters: what is the difference?
The model sets its parameters itself while it learns; people set hyperparameters to control how training runs. Google defines hyperparameters as “The variables that you or a hyperparameter tuning service adjust during successive runs of training a model.”5
Learn more on this site
- Size and other misused terms: AI terms people get wrong.
- How the numbers get set: How an AI is actually built.
- Changing them later: fine-tuning meaning.
- Who holds the weights: Open or closed.
- The basics: the AI glossary and AI explained for complete beginners.
- Related: What Is a Checkpoint in AI? Checkpoint Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What are parameters in an AI model?
- Limits a programmer writes by hand
- The settings you choose in an app, like length
- Numbers the model sets for itself while it learns
Show the answer
Answer: Numbers the model sets for itself while it learns. Parameters are learned, not written by a programmer; together they hold everything the model picked up in training.
2. What does “70B” in a model’s description mean?
- About 70 billion parameters
- It was released in 1970
- It costs 70 units to use
Show the answer
Answer: About 70 billion parameters. Models are often described by their parameter count, written with a B for billions.
3. Do more parameters always mean a better model?
- Yes, size is all that matters
- No, how much and how well it is trained matters too
- Yes, but only for image models
Show the answer
Answer: No, how much and how well it is trained matters too. In the 2022 study the smaller model, trained on more data, did better than the larger one.
Related terms: Model collapse · Open weights · Overfitting · Pretraining · Reinforcement learning · RLHF
- 1Google for Developers, Machine Learning Glossary: ML Fundamentals, undated glossary, read at source 1 Oct 2026a parameter is “The weights and biases that a model learns during training.” ↑
- 2NTIA (US Department of Commerce), Dual-Use Foundation Models with Widely Available Model Weights: Background, report published July 2024, read at source 1 Oct 2026“The values of these weights, and therefore the behavior of the model, are determined by training the model with numerous examples.” ↑
- 3Hoffmann et al., Training Compute-Optimal Large Language Models, arXiv, 29 Mar 2022, as cited on this site’s page on AI terms people get wrong, read at source 22 Sep 2026the model was trained with “the same compute budget as Gopher but with 70B parameters”. ↑athe model was trained with “the same compute budget as Gopher but with 70B parameters”. ↑b
- 4Hoffmann et al., Training Compute-Optimal Large Language Models, arXiv, 29 Mar 2022, read at source 22 Sep 2026 (via this site’s AI terms page)“Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks.” ↑a“Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks.” ↑b
- 5Google for Developers, Machine Learning Glossary: ML Fundamentals, read at source 1 Oct 2026hyperparameters are “The variables that you or a hyperparameter tuning service adjust during successive runs of training a model.” ↑ahyperparameters are “The variables that you or a hyperparameter tuning service adjust during successive runs of training a model.” ↑b
Hoffmann et al., Training Compute-Optimal Large Language Models, arXiv, 29 Mar 2022, read at source 22 Sep 2026.
The comparisons on this page (a mixing desk of sliders, engine size on a car) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK