What are model parameters — the 7B and 70B in the name

You've surely seen the odd model names: "Llama 70B," "Qwen 7B," "Mistral 8x7B." What's the B at the end? It's billions — billions of parameters. And here's the non-obvious payoff: once you get what that number is, a single name tells you whether a model will fit on your laptop and whether it's even worth chasing.

What a parameter is

Picture a giant console covered in knobs. Each knob nudges, just a little, what answer the model gives. A parameter is one of those knobs.

When a neural network is trained, it looks at examples millions of times and slowly tweaks all those knobs so the answers get better. After training, the settings are frozen. The finished model is that giant set of already-dialed-in knobs.

"7B" means 7 billion of those knobs. "70B" — seventy billion. Numbers a human can't picture, but routine for a computer.

Is more parameters always better?

Short answer: no — and that's the classic beginner mistake.

More parameters means more "capacity": the model is, in principle, able to catch finer patterns. But in practice what makes it smarter isn't only size — it's also the quality of the data it learned from and the polishing afterward. A small but well-trained model often beats a big, older one.

So don't pick a model by its B count like megapixels on a camera. Look at benchmarks — independent tests where models are compared fairly on real tasks. Size is a ceiling on ability, not a guarantee of it.

Why you'd want to know — the real practical payoff

Here's what made this worth reading. The parameter count directly tells you how much the model weighs and whether it fits your hardware.

A rough rule: in full format a model takes about 2 bytes per parameter. So:

• 7B is around 14 GB. Heavy, but doable on a good laptop.
• 70B is around 140 GB. Almost unrealistic at home; you need a serious server.

There's a trick — quantization: the numbers get squeezed, and that same 7B model fits into ~4 GB with barely any quality loss. So phrases like "this model runs on a laptop" usually mean the small, quantized versions.

The takeaway that stays with you: see "7B," think "light, fine at home"; see "70B+," think "this one's for the cloud." The number in the name is an instant hardware price-check.

Where you'll run into it

When you pick an open model to run yourself. When you compare options on OpenRouter or in a model catalog. When you read "a model with X billion parameters launched" — now you know whether it's huge or pocket-sized.

Are parameters and tokens the same thing?

No, and they're easy to mix up. Tokens are chunks of text you feed the model and pay for. Parameters are the model's internal "knobs," set once and for all during training. Tokens change with every request; parameters don't.

How many parameters does ChatGPT have?

For closed models like GPT or Claude, the exact numbers usually aren't disclosed — it's a trade secret. You mostly see the B count on open models you can download and run yourself.