I keep seeing model names with a number bolted on the end. 2B. 27B. A trillion. It looks like a score, or a price, or how many words the thing knows. It is none of those. It is a headcount of one specific thing.
That thing is called a parameter.
What they are counting
A parameter is a tiny number sitting inside the model. During training the model reads a mountain of text and keeps nudging those numbers until it gets better at guessing what comes next. When training is over, those numbers are the model. You are not downloading 27 billion words. You are loading about 27 billion little settings.
Two pictures help.
Think of a mixing board. Each parameter is a knob. A 2B model has about two billion knobs. A 27B model has twenty-seven billion. More knobs means more room to store patterns. It also means a heavier board. You do not get the extra knobs for free.
Or think of a library. The books are the text the model saw while it was learning. The parameters are the librarian’s habits: what usually follows what, which details matter, how to finish a sentence. A bigger parameter count is a librarian with more detailed habits. It is not a taller stack of books on your desk.
B is billion. T is trillion. A 2B model is small in anyone’s mouth. A 27B model is a serious piece of software. A trillion-parameter model is a different class of object.
Three numbers people mix up
| Term | What it is |
|---|---|
| Parameters | The knobs. That is the B in 27B. |
| Tokens | The chunks of text the model actually reads and writes. Not letters, and not always whole words. “Unbelievable” might be two or three tokens. “The” is usually one. A model trained on two trillion tokens chewed through that much text. It does not have that many knobs. |
| Context window | How many tokens it can keep in view at once. Desk space. You can have a small librarian and a wide desk, or a huge librarian and a narrow one. The B in 27B does not tell you the desk size. |
If you only remember one correction, remember those three are not interchangeable.
Why a “big” model sometimes fits a laptop
Those knobs take up memory. How much depends on how fat you store each one.
Full size is often 16 bits per knob. Quantization is writing the same knobs in shorthand. 4 bits instead of 16. Same count. Smaller file. You lose some fidelity, the way a compressed photo is still the photo. The machine requirement drops a lot.
That is how you get the headline that sounds like a magic trick: an 80 billion parameter model, quantized to 4 bits, on a laptop with 32 GB of RAM. The 80 billion did not shrink. Each number got cheaper to store. Whether that setup is comfortable depends on the rest of the stack, and it is tighter than the marketing slide. The point holds anyway. The parameter count and the machine it needs are not the same question.
What people mean by Small Language Model
There is no official cutoff. Nobody issued a standard. IBM talks about a few million to a few billion. Microsoft and others often treat under 10 billion as small. Some write-ups stretch higher. 2B is small. 27B is a gray zone. A trillion is not small.
You can look at the word four ways, and they do not always agree.
| Lens | What you are measuring |
|---|---|
| The number | How many knobs. Easy to say. Gets less useful as the years move and the tricks improve. |
| The bits | Same model, different memory bill. 80B at 16-bit wants a server. 80B at 4-bit starts a laptop conversation. |
| The machine | Phone, laptop, or a building full of GPUs. This is the one you actually feel. |
| The job | Narrower and more specialized, not just fewer knobs. |
I would not define small as “fits on a laptop.” That line is already moving. Quantization pulled it. The next trick, which I am going to write about separately, is only loading the parts of a huge model you need for the next word. If small means “runs on my machine,” next year’s compressed giant is small too, and the word stops earning its keep.
The useful claim hiding under SLM is this: it can run on-device, not in a data center. On-device still splits. Some of these fit a phone. Some need a laptop. Both are local. They are not the same size.
The question under the number
The B is a count of settings, not a grade. Bigger usually means more capacity and more hardware. It does not automatically mean better for the thing you are doing.
If someone says 27B, they mean knobs. If they say small, ask where it runs.