Model parameters: what 7B, 70B and trillions mean

Model parameters are the internal numbers of a neural network (weights and biases) that are tuned during training and store everything the model has learned. The number of parameters is often in the name: 7B means 7 billion.

In plain words

Imagine a huge mixing desk in a recording studio with billions of sliders. During training, each slider is nudged a little until the model "sounds" right. The parameters are the positions of all those sliders. The more there are, the finer you can tune the "sound" - but the desk gets huge and expensive.

How it works

1. Every connection between neurons has its own number - a weight.
2. During training, billions of these numbers are gradually changed so the model makes fewer mistakes.
3. After training, the numbers are saved to a file - and that file is the model.
4. The file size depends on the number of parameters and the precision of the numbers.
5. While running, all the parameters it needs must fit into memory - RAM or video memory.

Where you'll see it

Parameter counts appear in the names of open models (Llama 3 8B, Qwen 72B, Mistral 7B) and in news about new models. On-device AI uses small models - from a few hundred million to a few billion parameters.

Does it matter when buying

Yes, if you want to run a model on your laptop or phone. Roughly: a 7-billion-parameter model in compressed form takes a few gigabytes of memory. That is why for local AI I look at the amount of RAM. For cloud chatbots, the parameter count barely affects you.

Myths

• Myth: More parameters always means a smarter model. Reality: No. A new small model trained on better data often beats an old big one.
• Myth: Parameters are the number of facts the model knows. Reality: Knowledge is spread across all parameters, not stored one fact per parameter.

FAQ

What does B in a model name mean?

Billion - a billion parameters. 70B is 70 billion.

How many parameters do the biggest models have?

Companies often do not reveal exact figures, but estimates run to hundreds of billions and more.

How much memory does a 7B model need?

Compressed (quantised), a few gigabytes; uncompressed, noticeably more.

Related terms

Quantisation, Neural network, Open-weight models, On-device AI