What are parameters?
What the B in '8B' or '70B' means, and why a bigger number isn't always a better model.
You’ve probably seen model names like Llama 3 8B or 70B. That “B” stands for billions of parameters. It’s the most common way people describe how big a model is.
The numbers inside a model that get adjusted during training. Together, they hold everything the model has learned.
Picture a sound mixing desk with billions of tiny knobs. During training, each knob is turned a little at a time until the output sounds right. The final knob positions are the parameters.
That’s 70 billion numbers, each one tuned during training.
How parameters get their values
- 1Start random — every parameter begins as a meaningless number
- 2Guess — the model tries to predict the next piece of text
- 3Adjust — each wrong guess nudges the parameters slightly
- 4Repeat — billions of times, until the guesses get good
Small, medium and large models
- Small (a few billion)Fast, cheap, can run on a laptop
- Medium (tens of billions)A good balance
- Large (hundreds of billions+)Most capable, needs data centres
Why it matters
- Cost and speed. More parameters means more computing for every answer.
- Where it can run. Small models can run on your own computer, which keeps your data private.
- Capability. Bigger models usually handle harder tasks, but not always.
A common misunderstanding
“Bigger is always better.” Size helps, but so do the quality of the training data and the extra training afterwards. A newer, well-trained small model can beat an older, much bigger one.
Model makers sell the same family in several sizes. Anthropic’s Claude, for example, comes in several sizes, such as Haiku, Sonnet and Opus. The smaller ones are faster and cheaper; the larger ones are better at hard problems. Many apps send easy jobs to the small model and keep the large one for the rest.
Related buzzwords
LLM · Fine-tuning · Token