A 12B model that un-AI-ifies your drafts, offline and on your machine

What it does

It’s a 12B model, fine-tuned from Google’s gemma-4-12B, that takes AI-written drafts (emails, essays, reports, forum posts, in English and Chinese) and rewrites them to read like a person wrote them. It ships as a desktop app for macOS and Windows, or as a plain GGUF file for llama.cpp. Everything runs locally and works offline after the first model download.

The interesting bit

The quantization work is the sleeper hit. Instead of one standard llama-quantize pass, the 2-bit, Q3 and Q4_K_M files use sensitivity-aware bit allocation plus quantization-aware training and distillation from the full bf16 teacher. At 3.9 GB, a standard quant picks the same next token as bf16 70% of the time; theirs reaches 87% — and it’s still an ordinary GGUF any recent llama.cpp runs.

Key highlights

• 95% of English rewrites judged human by Originality.ai at its strictest setting — and no AI detector was used in training.
• 376 of 420 English rewrites had zero factual problems per a strict LLM judge; over 9 in 10 of the remaining fixes were a single word or phrase.
• Five model sizes, from Q8_0 (~12.7 GB) down to a 2-bit file at 3.9 GB, with the app suggesting one for machines down to 8 GB RAM.
• The app highlights new wording, strikes replaced wording, and lets you pin passages word-for-word with a “Create fact” feature.
• A dependency-free hz CLI rewrites whole Markdown or .docx files piece by piece, preserving headings, code blocks and tables — and reports any number that went missing or appeared.

Caveats

• It still changes details: the judge flagged 44 of 420 English rewrites (135 problems), e.g. turning “the remaining 37 complaints” into “the other 37% of complaints.” The README’s own advice: read the result before you send it.
• It’s a text-completion model, not a chat model — wiring it into chat endpoints will disappoint you.
• The app isn’t code-signed, so expect OS warnings on first launch.

Verdict

Worth a look if you write with AI assistance but don’t want your prose to smell like it — especially on privacy-sensitive text. Skip it if you need guaranteed factual fidelity; it’s a rewriter, not a copy editor.