The first story I generated while testing this tool was about a lighthouse keeper who finds a message in a bottle addressed to her. At 300 words it was genuinely good: a specific opening image, a keeper with a reason to care, a last line that landed. At 1,500 words, from the same premise, the keeper changed name twice, the bottle stopped mattering around the halfway mark, and the ending resolved a problem the story had not set up.
That gap is the single most useful thing to understand before you use any small language model for fiction, and it is not a flaw you work around by writing a better prompt. It is a property of the model. This guide covers what a 1B-parameter model does reliably, where it comes apart, how long your premise should be, and how to turn the output into something worth keeping, using the free AI story generator on this site.
TL;DR: The generator runs Gemma 3 1B on the Toolz server with eight genres and three lengths, roughly 300, 800 and 1,500 words. Below about 800 words it holds a scene, a voice and a shape. Above that it starts losing names, threads and the word target. Treat the 300-word setting as a finished sketch and the 1,500-word setting as raw material. Nothing you type is stored, logged, or used to train anything, which is the part that matters if you plan to publish the result.
What does a small model get right?
Four things, consistently.
It opens a scene. Give it a place and a person and it will put you somewhere specific in the first sentence, which is the hardest part of a blank page to get past. The prompt behind the tool asks for sensory detail and refuses the clichéd opening, so you rarely get "It was a dark and stormy night" or a character waking up.
It commits to a genre. Horror comes back with restraint rather than gore, children's stories come back with short sentences and a kind ending, and mystery comes back with a withheld fact. The eight genres are not a label on the same story; they change the vocabulary, the pacing and how much is explained.
It gives you something you would not have written. This is the real value and it is hard to measure. A premise about a message in a bottle produced a version where the bottle contained a shopping list, which was better than either idea I had. A generator is useful the way a writing prompt is useful: it moves you off the idea you were stuck on.
It varies. Sampling temperature is set high, so pressing Redo on the same premise gives a different story rather than a reworded one. Three runs of the same sentence are three drafts to choose between. That is a deliberate setting: a low temperature would make the tool feel broken, because every press of the button would return the same paragraph.
Where does it fall apart?
The failures cluster, and they all get worse with length.
Names drift. A character introduced as Maren becomes Mara and then Marin. The model has no memory of the text as a set of facts; it predicts what comes next from what came before, and over 1,500 words a name it saw a thousand tokens ago has less pull than one that sounds similar. Fixing this is a search and replace, which is why it is annoying rather than serious.
Threads get dropped. A gun on the mantelpiece in paragraph two does not go off. This is the one that costs real revision work, because the fix is not mechanical: you either cut the thread or write the payoff yourself.
The length target is approximate. Ask for 1,500 words and you may get 1,100. The model does not count; the length is a request in the prompt, not a constraint on generation. If a word count matters to you, paste the result into the word counter rather than trusting the setting.
Endings arrive early or not at all. Short stories tend to resolve; long ones tend to stop. A 1,500-word run that ends mid-scene is normal, and pressing Redo is usually faster than asking for a continuation.
Here is the pattern across the three lengths, from running the same four premises through each setting:
| Length | Typical time | Holds a voice | Keeps names straight | Resolves the ending | Use it as |
|---|---|---|---|---|---|
| Flash, ~300 words | 5 to 12 seconds | Yes | Yes | Usually | A finished sketch |
| Short, ~800 words | 15 to 30 seconds | Yes | Mostly | Often | A draft to revise |
| Long, ~1,500 words | 30 to 60 seconds | Drifts | No | Rarely | Raw material and ideas |
The honest reading of that table is that the flash setting is the one that produces something publishable with light editing, and the long setting is the one that produces the most material to work from. Neither is better; they are different jobs.
How much should you write in the premise?
More than a noun, less than an outline. One or two sentences is the band where the model has something to hold onto and still room to invent.
A bare noun gives you the average story about that noun. "A dragon" returns the dragon you have read a hundred times. Adding a want and an obstacle changes everything, because a want and an obstacle are what a story is: "a dragon who has lost her sense of smell and cannot find her own hoard" is a premise with a middle in it.
Going the other way, a full outline fights the model rather than steering it. If you specify five beats, a 300-word story has about 60 words for each, and what comes back reads like a summary of a story instead of a story. Past roughly three sentences of setup, you are better off generating the opening and writing the rest yourself.
The one instruction worth adding explicitly is a point of view, if you care about it. First person and third person both work; unspecified, the model picks, and it picks third person most of the time.
Which genre changes the output most?
Children's stories and horror, in opposite directions.
The children's setting is the most constrained: short sentences, concrete nouns, a problem solved by the character rather than by an adult, and an ending that does not frighten. It also has the fewest failure modes, because a 300-word bedtime story is well inside what the model holds together. If you are generating something to read aloud tonight, this is the setting that works best.
Horror is the most sensitive to the premise. Given a specific, mundane detail to be wrong about, it does restraint well. Given a generic premise, it reaches for blood. The difference between "a house that is haunted" and "a house where the upstairs hallway is three feet longer than it should be" is the difference between a story you have read and one you have not.
Fantasy and science fiction are the most verbose and the quickest to drift, because both invent proper nouns, and invented proper nouns are exactly what a small model forgets. Mystery is the most likely to drop its thread, since the withheld fact has to survive to the end. Romance and adventure sit in the middle. General fiction is the safest default and the least surprising.
What should you do with the draft?
A revision pass with four specific checks, in this order, is faster than reading for general quality.
Read the last paragraph first and ask whether it answers the first. If it does not, you have a sketch rather than a story, and the fix is usually to cut the last third and write a new ending.
Search for every proper noun and confirm it is spelled the same way throughout. This takes thirty seconds and catches the most embarrassing error the model makes.
List the objects and facts introduced in the first third, then check each one is either used or cut. This is the drift check, and it is where most of the real editing is.
Read one paragraph aloud. The model writes grammatical sentences that nobody would say, and the ear catches them when the eye does not.
What you should not do is publish it as it stands. Not because of the quality, which varies, but because of what the text is: a plausible continuation of a prompt, assembled from patterns. It can state something false with complete confidence, and it has no idea whether a phrase it produced is a line it absorbed from somewhere else.
Is it allowed where you want to submit it?
Often not, and the rules are getting more specific rather than less.
Many magazines now refuse machine-generated submissions outright. Clarkesworld closed submissions entirely in early 2023 under a flood of generated stories and has had a standing policy since. Competitions increasingly ask for a declaration. The Authors Guild has published guidance on disclosure for its members, and several publishers now require it in contracts.
The practical rule: a generated draft you rewrite is your work, a generated story you submit is not yours to submit, and anything in between needs the venue's own rules rather than your judgement. If you are a student, your institution's policy governs, and "I only used it for the first paragraph" has not been a successful defence.
None of that applies to the uses this tool is good for: getting unstuck, generating twenty premises and keeping one, writing a bedtime story, or drafting something nobody but you will read.
How does this compare to a large commercial model?
Differently enough that the comparison is not about quality.
| Gemma 3 1B here | A large commercial model | |
|---|---|---|
| Plot consistency over 1,500 words | Drifts | Holds |
| Prose at the sentence level | Competent | Better |
| Speed for flash fiction | 5 to 12 seconds | Similar |
| Cost | Free, no signup | Free tier or subscription |
| What happens to your text | Discarded after the reply | Depends on the provider and your settings |
| Works on an unfinished idea you care about | Yes | Read the terms first |
The size difference is real. Gemma 3 at 1B parameters is roughly three orders of magnitude smaller than a frontier model, and plot coherence is one of the capabilities that scales with size. If you need a 5,000-word story that holds together, this is not the tool.
The reason to use a small model anyway is the fourth and fifth rows. Your premise goes to our server over HTTPS, the server builds the prompt, Ollama runs the model, and the story streams back. The text is used to produce the reply and then discarded. It is not stored, not logged, not used for training, and not sent to any third-party AI company. For an idea you intend to publish under your own name, that is worth more than better prose, because the alternative is pasting unpublished work into a service whose retention terms you have to go and read.
The rate limit is about twenty stories every ten minutes, which is generous enough that Redo is a normal part of using it.
Where to go next
If the story is not the format you wanted, the same model sits behind the poem generator and the YouTube script writer, both with prompts built for their shape rather than for narrative. Everything built on this model is listed on the AI tools hub.
And if what you need is the opposite of generation, the honest answer is a prompt list and an hour. A generator is good at the first sentence, and the first sentence was never the hard part of finishing something.
