Model parameters: what the “B” in 7B or 70B really means
Parameters are the internal numbers an AI model adjusts during training, like billions of tiny tuning knobs. Their count tells you how big a model is and roughly how much memory it needs to run, but not, on its own, how good it is.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
A simple definition
A modern AI model is a neural network: a massive chain of calculations in which each connection carries a weight, a number. These weights, plus a few other values, are the parameters. They start out random. During training, the model reads data, makes mistakes, and each mistake is used to adjust its parameters slightly. By the end, the parameters are where everything the model learned “lives”: grammar, facts, styles, reasoning patterns.
A historical benchmark: in 2020, the paper introducing GPT-3 described a model with 175 billion parameters, presented at the time as ten times larger than any previous non-sparse language model.
A concrete example
You want to run an open-weight model on your computer. Its model card says 8 billion parameters. If each parameter is stored in 16 bits (2 bytes), you need about 16 GB of memory just to load it. With 4-bit quantization, that drops to roughly 4 to 5 GB, at the cost of a slight loss in precision. So the parameter count is first and foremost a way to tell whether a model will fit on your hardware. Our open-source models page shows the size of the most downloaded models.
Why it matters
More parameters generally let a model store more knowledge and nuance, but make it more expensive to train and run. That’s why labs look for trade-offs: well-trained small language models, or mixture-of-experts designs where only part of the parameters work on each token. In that case you’ll often see two figures: total parameters and active parameters.
Common myths
- “More parameters means a better model.” Not necessarily. Data quality, training length and post-training tuning matter just as much. A recent smaller model can beat an older, bigger one.
- “Parameters store knowledge like an encyclopedia.” Knowledge is spread across billions of numbers; you can’t go in and edit one entry by hand.
- “Fine-tuning rewrites the whole model.” Methods like LoRA train only a small fraction of extra parameters.
Related terms: neural network, quantization, mixture of experts, small language model, open-weight model.
Related terms
- Neural network · Software made of many small connected computing units that learns by adjusting the strength of the links between them.
- Training data · The examples (text, images, audio and more) an AI model learns from; they shape what the model knows and where it goes wrong.
- Mixture of experts (MoE) · An architecture in which a model holds many expert sub-networks but only switches on a few of them for each token it processes.
- Quantization · A technique that stores a model's weights with fewer bits (8, 4 or even fewer) so it needs far less memory to run.
- Small language model (SLM) · A compact language model, light enough to run on a laptop or phone, that performs well on focused tasks rather than everything at once.
- Open-weights model · An AI model whose trained parameters (the weights) can be downloaded, so you can run it on your own hardware.
Try it
Full glossary
Frequently asked questions
What does “7B” in a model's name mean?
B stands for billion. A “7B” model has roughly 7 billion parameters. It is a size indicator, useful for estimating the memory needed, not a quality score.
Are parameters the same as training data?
No. Training data is the text the model reads; parameters are what it retains from it, stored as numbers. A model doesn't hold a copy of its data, even though it can sometimes reproduce passages.
Why don't some AI companies disclose parameter counts?
Many labs keep this confidential for their closed models, for competitive reasons. Open-weight models usually state it, since their files are public anyway.
Does “parameter” have another meaning?
Yes, which causes confusion: in an API, the settings of a request, such as temperature or maximum output length, are also called parameters. You choose those; the model's own parameters are fixed by training.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.