A 12-billion-parameter model fine-tuned on paired drafts — stiff on one side, natural on the other. Paste your text, get back the same meaning in prose that reads like a person actually wrote it.
How it works
Rather than being told what good writing looks like, the model learned from thousands of paired examples — the same content written stiffly and written naturally — and inferred the mapping between them.
Anything up to 3,000 words. Reports, emails, posts, documentation — it's trained on general prose, not one narrow genre.
Sentence rhythm, word choice and structure all shift. The meaning, the facts and roughly the length stay where they were.
Generation is sampled, so every run is a different take. Not quite right? Re-roll and get another.
The tool
Paste text on the left. The rewrite streams in on the right as it generates.
Ctrl + Enter to run · first request after idle takes ~30s to warm up
FAQ
It shouldn't. The model is trained to preserve meaning and facts while changing how they're expressed. That said, it samples at a high temperature, so read the output before using it — particularly numbers, names and dates, which can occasionally drift. Re-roll if something looks off.
Generation is sampled rather than deterministic, which is what lets the output read naturally instead of formulaically. The trade-off is that identical input produces different output each time. Re-roll is there to take advantage of that.
The GPU scales to zero when idle, so nothing is running — and nothing is being billed — between requests. The first request after a quiet period loads about 10GB of model weights into VRAM, which takes roughly half a minute. Requests after that are fast.
3,000 words per request, bounded by the model's 8K-token context window. For longer documents, split them by section and run each separately.
It's sent to the inference endpoint, rewritten, and streamed back. It isn't stored or used for training.