Humanizer 12B: A Local AI-Text Rewriter, and What Its Own Numbers Actually Say

Timeline

1. September 11, 2026

   The humanizer repository appears on Hugging Face. The card describes a 12B model, fine-tuned from google/gemma-4-12B and released under Apache 2.0, that rewrites AI-written drafts in English and Chinese so they read as if a person wrote them, while keeping every number, unit, date, name and quote.

   jialinyyzz/humanizer model card

2. October 2, 2026

   The author runs Originality.ai at its strictest setting on 210 English rewrites made with the bf16 weights: 11 are flagged as AI, against 26 for the previous release. The card stresses that no AI detector was used anywhere in training.

   humanizer model card, Evaluation details

3. October 4, 2026

   From this date every GGUF file stores the recommended sampling settings in its metadata, and carries a chat template that builds the model’s prompt from the last user message, so chat front ends such as llama-server’s chat endpoint can drive a model that is really a text-completion tool.

   humanizer model card, Prompt format

What it is, and what it is not

Most tools sold as “AI humanizers” are web forms. This one is a model you download. The card describes a 12B-parameter rewriter that takes a draft written by a language model and rewrites it so it reads like a person wrote it, in English and in Chinese. The design goal is stated in a single constraint: keep every number, unit, date, name and quotation, and add nothing. It runs locally, there is an app for Macs with Apple silicon and for Windows, and the card also documents llama.cpp, MLX, transformers, vLLM and Ollama routes. The model is fine-tuned from google/gemma-4-12B, and both weights and code are under Apache 2.0.

It is not a chatbot. The card is emphatic that this is text completion with no system prompt and no turn markers: you send a fixed instruction, a blank line, your draft, a separator and the words “Rewritten:”, and let the model continue. The card tells you to stop on end-of-sequence only, never on a stop string, and to use temperature 1.0 and top-p 0.95 with top-k and min-p switched off, because llama.cpp’s defaults and the bundled generation config would otherwise interfere. That fussiness is a sign of a tool built for a single job.

How it was made, according to the author

The training story on the card is detailed enough to be checkable, which is rare. The first stage was supervised fine-tuning on 28,598 pairs, each a language-model-written draft paired with a real human original; the human side is described as paper abstracts, government reports, student essays, company and mailing-list email, Reddit, Hacker News and Zhihu. The second stage was preference training on 3,918 pairs chosen only for fact fidelity and copying, judged by an LLM. The third was reinforcement learning in three rounds totaling 500 steps, with a strict single-vote fact judge in the first 200 and two rounds of 150 after that; each step used 16 drafts with 8 samples each.

The reward combined an LLM judge, which read each rewrite against its draft and penalized severe errors, invented content, changed meaning and dropped formatting, with a penalty for copying the draft’s five-grams and sentence skeletons. The card says the data behind every row used for the second release is published in a companion dataset. And it makes a claim that matters for how to read the headline: no AI detector was used anywhere in training, not as a reward, a filter or a way to choose a checkpoint.

The headline: 95% judged human

The card’s headline says 95 percent of 210 English rewrites were judged human by Originality.ai at its strictest setting on October 2: 11 were flagged as AI, against 26 for the previous release, a difference the card reports as statistically significant on paired data. It also compares against a public baseline, a humanizer skill with about 53 thousand GitHub stars, applied by Claude Sonnet to 60 of the same drafts, where all 60 were flagged. The new model, on the same 60, had 4 flagged.

Three reasons to hold that number loosely, all of which the card itself supplies. It is one detector on one date, and the card says detectors change and nothing guarantees any detector outcome. The evaluation drafts come from the author’s held-out set, written by three frontier models in roughly equal shares, so they say little about, for example, a blog post a human started and an AI polished. And the most templated genres remain hardest: the card’s breakdown shows emoji-and-hashtag social posts at 3 flagged of 16 and formal policy memos at 2 of 13, while emails, blog posts, Reddit posts and product reviews had none flagged.

The numbers the card puts in the open

The more useful figures are the ones about facts. A strict LLM judge found no factual problem in 376 of 420 English rewrites; the other 44 were flagged, with 135 problems in all, 125 of them a single word or phrase. The card’s example is a draft saying “The remaining 37 complaints” that came out as “The other 37% of complaints”. For Chinese the picture is weaker: no factual problem in 149 of 204 rewrites. The previous release scored 369 of 420 in English and 135 of 204 in Chinese, so the direction of travel is good and the absolute level is not perfect.

The card also lists failure modes that are not about facts. 28 of 420 English outputs dropped a format element such as a list or heading. Short, informal English drafts with technical jargon occasionally come back written entirely in Chinese, a failure the app and the command-line tool detect and retry. In casual genres it sometimes adds slang or profanity that was not in the draft. And the before-and-after examples on the card are picks: the best of eight samples among those the fact judge passed.

There is even an audit of the judge. A second judge re-read 150 rewrites the first had flagged as serious; it agreed the problem was serious in 99 cases, rated it minor in 41 and found the fact unchanged in 10. The authors’ conclusion is that some flags are harmless rewording, and they still tell you to read numbers, dates and names before you send.

Why that matters more than the 95%

A tool that is mostly right about numbers will, over enough documents, change one. The card says so: read the result before sending it. That is the real cost of using it, and the reason the command-line tool’s habit of flagging any piece where a number went missing is worth more than a few points of detector score. The speed numbers help set expectations: on an M5 Max, the app’s llama.cpp setup runs at roughly 36 to 38 tokens a second, so a hundred-word email takes about 3.6 seconds and a Chinese email of about 300 characters about 8.5.

The card is also explicit about the ethics. It describes the tool as a writing tool for your own drafts and says that where a school, employer or publication has rules on AI assistance, you should follow them. We agree, and would go further: a rewrite that merely evades a detector has not made the underlying draft better. The version of this tool worth wanting is the one that makes a stiff draft read like a person, with its facts intact, on your own machine.

If you only want a hosted option

Our directory lists Grammarly AI Humanizer, a free hosted tool that turns AI text into natural writing while keeping the meaning. It is a different trade: no setup and no local hardware, but your text goes to a service. For a decision about which of this model’s files fits your machine, see which humanizer file for 8, 16 or 32 GB.

Our take

This is a well-documented release with the weakness of the genre: the claim most people will repeat is the one least connected to the quality that matters. Use the fact-check figures, not the detector figure, to decide whether it fits your work. If you write documents full of numbers, budget time to proofread every rewrite. If you write essays and emails with few hard facts, the risk is lower and the benefit is a more natural voice.

Keep reading on Front

Every post, newest first, is on the Front blog index.

• Humanizer 12B: which GGUF fits your memory: how much memory each GGUF file of the same model needs.
• Ling 3.1 Flash: what it takes to run: the hardware needed to run Ling 3.1 Flash.
• Grok 4.8: what xAI has actually said: xAI’s Grok 4.5 to 4.7 releases and what has been said about 4.8.
• Google Argon timeline: the dated record of Google’s Gemini 4 Argon launch.

FAQ

Update log

• October 17, 2026 — First published.

Sources

• jialinyyzz/humanizer model card
• Grammarly AI Humanizer