Three models, sixteen to a hundred and thirty-seven megabytes. Two of them only know children's stories and write them beautifully. The third knows everything and gets it wrong.
cooking · work in progressThey finish your sentence rather than answer you — they continue text, which is the older and plainer thing a language model does. Neither is looking anything up.
For most of their history these things did one job: you gave them the start of something and they carried on writing. Asking one a question is recent, and it is a separate skill that has to be taught on top — which is why the small models here can do the first and not the second. Both paths are below. Try the same thought in each.
The older job, and the plainer one. It takes what you wrote and keeps going. Every model here can do it.
The newer job. It treats what you wrote as a question and replies. Only a model taught to hold a conversation can do this.
Three attempts at the same thing
Models do not read letters or words. They read tokens — chunks of roughly three or four characters, so a short common word is usually one token and a longer or unusual one is two or three. Every attempt above reports how many went in and how many came out, because that is the unit everything is counted in: what you are billed, what a rate limit allows, and how much a model can hold at once.
Prices are quoted per million tokens, which sounds enormous and is not. A short answer here is well under a hundred tokens, so a million of them is only some thousands of answers. Change the dial, change the mode, change the model, and watch the count move — asking a question costs far more than continuing a sentence, because the conversation format wraps your words in instructions the model has to read first.
At every word the model produces a ranked list of candidates with a confidence attached to each. The dial decides how strictly it obeys that ranking.
All the way down, it always takes the single likeliest word. That makes it repeat itself — run it three times and you get the same sentence three times, because nothing in the process is random. Turn it up and less likely words get a real chance, so the three attempts diverge. Turn it up far enough and words that were never plausible start winning, which is where sentences come apart.
The useful thing to notice is that the small model falls apart much sooner. It has less confidence to spare, so the same amount of added risk destroys it while the larger one is still merely loosening up.
| Writer | Size | Download | Parameters | Per attempt, in a browser |
|---|---|---|---|---|
| Storyteller | XS | 15.7 MB | 3.7 million | ~0.9 s |
| Storyteller, grown | M | 110 MB | 110 million | ~1.4 s |
| SmolLM2 | M | 137 MB | 135 million | ~2.4 s |
The size labels are about what the download costs you, not about the model: XS is under twenty-five megabytes, S under a hundred, M under four hundred.
Put that beside the other tiles on this wall and something stands out. A privacy filter does a genuinely useful job at 17 MB. Depth recovers the geometry of a photograph at 19 MB. A voice is understandable at 24 MB. But to get a model that merely writes a grammatical English sentence you need 137 MB — and it will still get its facts wrong.
Perception is cheap. Language is expensive. That is the through-line of the whole series, and this is the tile where the price tag is most obvious.
Every one of these was timed in a browser on the same machine, writing the same forty-eight words. If speed followed size, the table would descend neatly. It does not.
| Model | Parameters | Where it ran | Words a second |
|---|---|---|---|
| Storyteller | 3.7 million | processor | 122–774 |
| Storyteller, grown | 15 million | processor | 11 |
| a 33-million story model | 33 million | processor | 4 |
| a 110-million story model | 110 million | processor | 4 |
| SmolLM2 | 135 million | processor | 12 |
| SmolLM2 | 135 million | graphics chip | 60 |
| Storyteller, grown | 15 million | graphics chip | 98 |
The largest model on the list beats two smaller ones by three times. A 33-million model and a 110-million one run at exactly the same speed. Both of those are packaging, not arithmetic: a model can be published in a form that remembers what it has already read, or in one that re-reads the whole sentence before choosing each next word. The second is far easier to publish and quietly costs you most of your speed.
Then the last two rows: the same models on the graphics chip, five to nine times faster again. Which is why this page uses it when your browser offers it, and why the number of parameters — the figure everyone quotes — tells you remarkably little about what an experience will feel like.
The Storyteller is thirty-seven times smaller than SmolLM2 and writes English just as cleanly — complete sentences, subjects agreeing with verbs, a story that follows from its own first line. It manages that on 3.7 million parameters because it was only ever taught one thing: simple children's stories, in the vocabulary a four-year-old uses.
Ask it anything else and it does not refuse or stumble. It simply carries on telling you a story, because a story is the only shape it has. Give it The capital of France is and you will get a child thanking a man for his help.
So the honest ladder is not one line from bad to good. Below about a megabyte there is no English at all. A few megabytes buys fluent English about one narrow thing. Going general — writing about anything at all — is what costs the other hundred and twenty megabytes, and it still gets its facts wrong. Narrow is cheap. Broad is expensive.
While building this I tried a third model — four times larger than the biggest one here
— and it produced also the same as well as so forth to so forth to so forth
. An
easy conclusion: bigger is not better, the ladder is broken, leave it out.
That conclusion was wrong. The model expects to be addressed in a particular format, as a
conversation rather than as a half-finished sentence. Asked the way it expects, the same file
answered Paris
immediately and correctly.
It is left off the ladder on size: 545 MB against SmolLM2's 137, for an answer that is shorter rather than better. But the near-miss is worth recording: most reports that a small model is useless are really reports about how it was asked.
The Storyteller is part of this site and arrives with the page. The larger one downloads from a public model hub the first time you pick it, then your browser keeps it.
What you type is never uploaded. There is no server here to receive it, and the writing happens on your own machine.