AI Paraphraser
Paraphrase text four ways: formal, casual, shorter, or longer, with a small language model that does the rewriting on your own graphics card.
Rewrite
The rewrite lands here, one row per pass.
The rewrite stopped
Canceled.
How to use
- Type or paste into Text to rewrite. One awkward sentence is a perfectly good input, and so is a three-paragraph draft.
- Choose a card under Rewrite as. Formal is preselected, and the four cards are four different instructions to the model rather than positions on a slider.
- Press Rewrite. The first press fetches the model, at the size the badge states, and the bar counts that download in bytes before it counts the passes. Cancel sits beside the bar throughout.
- Read the rows in the Rewrite box, press Rewrite again if the wording missed, pick the take you like from the chips, then press Copy rewrite.
How the rewriting works
Rewriting your own sentence is oddly hard. You already know what it means, so every attempt comes back the same shape as the first one. A paraphrasing tool helps here for a dull reason: the model has no attachment to your draft. Hand it the line, choose a target, and read a version somebody else would have written.
Rewrite as holds four options and opens on Formal. Formal asks for complete sentences, precise vocabulary, and no contractions. Casual asks for the reverse, simple words and a friendly tone. Shorten is told to cut extra words and unnecessary detail while keeping every main idea. Expand is told to add examples, clarifications, and context to something too terse to stand on its own.
A small instruction-tuned language model does the writing, on your graphics card through WebGPU. The badge above the panes says how much the first press fetches and whether the model is already on this device; after that one download the page rewrites with the network off. Chrome, Edge, Safari 26, and current Firefox can reach a GPU. Where the browser or the card falls short, Rewrite stays greyed out and the badge names which of the two is missing.
Longer text is never fed in whole. The planner cuts at paragraph breaks first, then at sentence ends, aiming for pieces of about 150 words, and every piece is a slice of exactly what you typed, so 3.5 p.m. and 12.4% reach the model as written. Each piece is a separate pass with its own row in the Rewrite box, filling in while you read, and a small mark in your text shows where each pass begins. Blank lines between your paragraphs are recorded before the split and put back afterwards.
Sampling is switched on at temperature 0.7. Run the same paragraph twice and the two rewrites will part company inside the first line, which is the point: the page keeps your last three takes side by side, so press a few times and copy the one that sounds like you. Read it back before it goes anywhere, because 0.6 billion parameters is small enough to drop a qualifier without noticing.
One blunt line, made formal
Start with a Slack-grade sentence: “Can't do the Tuesday call, move it.” Leave Rewrite as on Formal and press Rewrite. Seven words in means the token budget lands on its floor of 64, which is still several times the room the answer needs, so nothing gets truncated.
Back comes one polite sentence, along the lines of “I am unable to attend Tuesday's call and would like to reschedule it.” Press Rewrite again and the wording shifts, because the decoder samples rather than always taking its top guess; the new version arrives as Take 2, with the first still a press away. Switch to Expand and you will get a reason for the change invented for you, which is precisely the moment to edit instead of copy.
Good to know
- Chunks are rewritten independently, so the model never sees the piece before the one it is working on. That is where meaning drifts: a pronoun loses the thing it referred to, a term you defined in the opening comes back undefined, and a comparison loses its other half. Passes that share a paragraph are named under the result, and the marks in your text show where each one starts, so read closely there.
Frequently asked questions
Why is the Rewrite button greyed out?
Either the box is empty, or your browser could not hand the page a GPU. There is no processor fallback here on purpose, since a weaker model tends to invent rather than rephrase. Chrome, Edge, Safari 26, and recent Firefox all work. Some GPUs pass the WebGPU check but lack 16-bit shader support, which this model needs, and the badge says which of the two cases you hit.
What is the real difference between Shorten and Expand?
The instruction, and the room the model is given. Shorten gets a budget of roughly 1.2 times the input length and is asked to strip padding while keeping every idea. Expand gets about three times the input and is asked for examples and extra context. Both are capped at 768 new tokens per chunk, so a very long piece cannot run away.
Will a rewrite from here get past an AI detector?
I would not build anything on that assumption. The output is written by a language model, so a detector looking for machine text has every reason to flag it. Disguising authorship is the wrong job for this page anyway. What it is good at is handing you a second phrasing of your own sentence to choose from.
Why does the same paragraph come out different every time?
The decoder picks each word from a probability distribution instead of always taking the most likely candidate, at temperature 0.7 with top-p 0.8. Two runs diverge almost immediately. Treat it as a drafting feature: three presses give you three takes, kept as chips so you can flip between them.
How much text can I put in?
The box has no ceiling. What grows is the work. Twelve hundred words gets planned into at least eight passes and usually nine, since a pass ends at the last sentence that fits in 150 words. Each one is a separate trip through the model, and the bar counts them off, Rewriting 4 of 9, so you can see how far along it is.
Do my paragraph breaks survive?
They do. The planner notes which pass ended each paragraph before any rewriting starts, and Copy rewrite puts a blank line back at exactly those positions. Passes inside one paragraph rejoin with a single space, line breaks the model writes inside a pass are kept, and stray spaces before commas and full stops get tidied up. If a pass comes back empty, its part of your text stays as you wrote it, so no paragraph goes missing.
Why is the first rewrite so much slower than the rest?
Two things have to finish before a single word appears. The model downloads, at the size the badge states, then it is compiled into shaders for your particular card. Neither happens again. From your next visit the file comes from cache and the page keeps working with the network off.
Which model does the rewriting?
Qwen3-0.6B, an instruction-tuned model from the Qwen team published under Apache-2.0, run with 4-bit weights and 16-bit arithmetic through Transformers.js. It runs in non-thinking mode with the sampling its model card recommends for that mode. There is no processor path: measured without a GPU it wrote about half a token a second, far too slow for a paragraph.
Can I rewrite one sentence in the middle of a paragraph?
Not from the box as it stands. Paste that sentence on its own, rewrite it, and drop it back into your document. Paste the whole paragraph and every sentence in it is fair game, so the model will cheerfully reword the parts you were happy with.
Missing something in AI Paraphraser? Suggest a feature →
Tell someone who needs this.