AI Paraphraser
Rewrite a passage in a different tone, shorter, or more casual. Flan-T5 runs in your browser — no API key, no upload.
Why a Paraphraser That Runs Locally
The standard online paraphraser workflow involves pasting your text into someone else's web app and trusting that they will not retain it. For draft content, internal communication, or anything else you would prefer not to send to a third party, that trust is hard to verify. This tool runs Google's Flan-T5-small model in your browser via transformers.js. The text you paste stays on your machine. The model is roughly 80 MB on disk and downloads once on first use, after which everything runs locally. Flan-T5 is an instruction-tuned model — it responds to natural-language prompts like "rewrite this in a more formal tone" or "summarize this in one sentence." That makes it useful for paraphrasing across tone, length, and style. The small variant is fast enough to run interactively even without WebGPU, and the Apache 2.0 license released by Google makes it usable in commercial work with no strings.
How the Paraphraser Works
The first visit downloads the Flan-T5-small model — about 80 MB, cached in your browser thereafter. Paste your input text, pick a rewrite style from the dropdown (Formal, Casual, Shorter, Longer, Simpler, or Custom), and click Rewrite. Internally the tool prepends an instruction prefix to your text — for example, "Rewrite the following text in a more formal tone:" — and feeds the combined prompt to Flan-T5. The model returns one paraphrased candidate, displayed in the output panel with a copy button. The Custom style lets you write your own instruction prefix, which is the most powerful mode — try things like "Rewrite for a 10-year-old reader," "Convert to bulleted list," or "Translate to passive voice." The output is bounded to about 256 tokens (around 200 words) per pass; for longer inputs, paraphrase in sections. Flan-T5-small is a small model by current standards — its rewrites are coherent but less polished than a hosted GPT-class model. The trade-off is that nothing leaves your machine and there is no per-request cost.
Frequently Asked Questions
What is the underlying model behind the paraphraser?+
Google's Flan-T5-small, served via the Xenova ONNX port for transformers.js. Flan-T5 is an instruction-tuned variant of T5, released by Google under the Apache 2.0 license. The small variant is approximately 80 MB compressed.
Where does my draft text go when I click Rewrite?+
Nowhere off your device. After the model downloads on first use, every paraphrase runs entirely in your browser. The text you input never leaves your machine, and there is no telemetry on the content you process.
How is paraphrasing different from summarizing?+
Summarizing produces a shorter version that preserves the main points but drops detail. Paraphrasing produces a roughly same-length rewrite in different words — useful for changing tone, simplifying vocabulary, or avoiding direct copying. The Summarizer tool is the right pick for the first job, this tool for the second.
What rewrite styles are available?+
Six built-in styles: Formal, Casual, Shorter, Longer, Simpler, and Custom. The Custom style lets you write any instruction you want — Flan-T5 is instruction-tuned, so it responds to natural prompts like 'Rewrite this for a 10-year-old reader' or 'Convert to bulleted list.'
What is the maximum passage size per rewrite?+
Up to about 512 tokens per pass — roughly 400 words. For longer text, paraphrase in sections. The output is capped at about 256 tokens (around 200 words) to keep latency reasonable.
Why are the rewrites sometimes very similar to the input?+
Flan-T5-small is a smaller model than what hosted services use, and on short or already-clean inputs it sometimes returns nearly-identical output. Try a more specific Custom instruction (for example, 'Rewrite using only one-syllable words' or 'Convert all sentences to questions') to force a stronger rewrite.
Is the output safe to ship in client deliverables or academic work?+
Yes — the model is Apache 2.0 licensed, so commercial use is permitted. As with any AI tool, treat the output as a starting draft to review and edit, not a finished deliverable. Always disclose AI assistance where institutional policies require it.
Why does the model download every time I clear my browser cache?+
Model files are cached in IndexedDB. Clearing site data, using private browsing mode, or switching browsers triggers a fresh download. Browsers occasionally evict large IndexedDB stores under storage pressure, which also forces a re-download. The model is served from the Hugging Face CDN at full speed.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service