Also called: weights · Last checked: 7 November 2026
In short: A parameter is one of the adjustable numbers inside an AI model that store what it learned in training. Models are often described by how many they have, from millions in small models to a trillion or more in the largest.
Table of Contents
How it works
- Start. Before training, a model’s parameters are set to random values.
- Adjust. During training, each parameter is nudged up or down to reduce errors on the examples.
- Save. After training, the final values are stored as the model’s weights and used to answer new questions.
Where you’ll meet it
Model names such as Llama 3 70B or Mistral 7B give the parameter count in billions. The makers of ChatGPT, Gemini and Claude do not publish counts for their main chat models.
What it means for your business
- Size is a rough guide. More parameters usually mean more ability, but also higher cost and slower answers.
- Smaller can fit the task. A model with a few billion parameters can run on your own hardware for narrow jobs.
- Open weights give control. When the parameters are published, you can run and adapt the model yourself.
Watch out for
- Count is no proof of quality. Training data and tuning matter as much as size, so test models on your own tasks.
- Hardware needs. Large models need expensive chips with plenty of memory to run.
Example
Technology Innovation Institute. In September 2023 Abu Dhabi’s Technology Innovation Institute (TII) released Falcon 180B, a model with 180 billion parameters trained on 3.5 trillion tokens. TII called it the world’s most powerful open language model. Anyone could download its weights, though the licence restricted some commercial uses, such as offering it as a hosted service.
Parameter vs token
| Parameter | Token | |
|---|---|---|
| What it is | A number inside the model that stores learning | A piece of text the model reads or writes |
| When it matters | Set during training | Counted every time you use the model |
| What it tells you | Model size | Usage and cost |
Frequently asked questions
Does a model with more parameters always give better answers?
No. More parameters help, but training data, tuning and design matter too. Newer small models often beat older large ones on everyday business tasks.
Are parameters and weights the same thing?
Mostly, yes. Weights are the main type of parameter, and people use the two words interchangeably. “Open weights” means the trained numbers are published for anyone to download.
Why do some AI firms keep parameter counts secret?
The count hints at training cost and design, which rivals could use. Firms also prefer buyers to judge models on results rather than on size.