Parameters (Model Size)
Parameters are the many individual numbers inside an AI model that get adjusted during learning and that store the model's entire knowledge. Their count is considered a common measure of a model's size – you encounter it in figures like "7 billion parameters".
At its core, an AI model consists of a very long list of numbers. These numbers are called parameters. They are not set by humans but are adjusted automatically while the model learns from example data. Afterward, they remain fixed. Everything the model can do is contained in these numbers. How many there are is usually stated as a size figure: a model with 7 billion parameters, one with 400 billion. In product names, this often appears as an abbreviation, such as "8B" for eight billion.
Why the number matters
The parameter count is the simplest metric for comparing models. Roughly speaking: more parameters mean more capacity, i.e. more room for knowledge and patterns. That's why, for years, models were made mainly bigger.
But the number also determines cost and technical requirements. Every parameter needs storage space. A model with 70 billion parameters doesn't fit on a laptop; it needs expensive specialized chips in a data center. Small models with just a few billion parameters, by contrast, run directly on a phone.
Important: bigger is not automatically better. A smaller model with very good training data regularly outperforms a larger one with poor data. The parameter count says something about the effort involved, not directly about quality.
How parameters arise and work
You can think of parameters as the knobs on a huge mixing console. Each knob influences a tiny piece of the calculation that turns an input into an output. At the start, all the knobs are set randomly. The model then produces nonsense.
During training, it's given examples and is supposed to, for instance, predict the next word of a sentence. If it gets it wrong, all the knobs are nudged a tiny bit in the direction that reduces the error. This is repeated billions of times. After weeks on thousands of chips, the knobs are set in such a way that sensible answers come out.
No single parameter stands for a particular fact. The knowledge is distributed across a great many numbers. That's precisely why no one can simply look up why a model gave a particular answer.
Two details are often confused. First: a model can have many parameters and still compute quickly if only part of them is used per query. This is referred to as total parameters and active parameters. Second: exactly how precisely each number is stored is a separate question. If you store it more coarsely, the model shrinks in memory without any parameters being removed.
Where the term appears
In news about new AI models, the parameter count is almost always the first metric mentioned. For freely downloadable models, it appears directly in the name, for example in the variants of the Llama or Qwen series. Anyone wanting to run such a model themselves checks this number first to see whether their own hardware is sufficient.
For commercial services like ChatGPT or Gemini, providers now mostly no longer state the number. It's treated as a trade secret, and companies prefer to highlight benchmark results instead. If estimates circulate anyway, they are rumors, not confirmed figures.
The number also plays a role in discussions about regulation and energy demand. Large models mean large data centers, high electricity bills, and strong demand for AI chips. When stock market reports write about the demand for such chips, growing model size is one of the reasons behind it.