What is humanizer (12B)?
humanizer is a 12-billion-parameter model fine-tuned from Google's Gemma 4 12B by an independent developer who publishes as jialinyyzz on Hugging Face and sgaofen on GitHub. It does one job: you paste an AI-written draft (an email, essay, report or forum post, in English or Chinese) and it rewrites it so it reads as if a person wrote it, while being trained to keep every number, unit, date, name and quotation. The first release appeared in September 2026; the current v2 weights were evaluated on 2 October.
It is a text-completion model, not a chat model, with an 8,192-token window for instruction, draft and rewrite together. It ships as safetensors and as six GGUF files from 3.9 GB (2-bit) to 23.8 GB (bf16), several of them quantisation-aware trained, plus a free offline app for Mac and Windows and a command-line tool that rewrites .md, .txt and .docx files piece by piece.
Training used 28,598 pairs of AI draft and real human original, 3,918 preference pairs and 500 steps of reinforcement learning scored by an LLM fact judge. The author says no AI detector was used anywhere in training, and has published every training row. Weights and code are under Apache 2.0.
Is humanizer (12B) worth it?
humanizer's selling point is that it makes AI text pass AI detectors, and the card says so plainly: on 210 held-out English drafts, Originality.ai at its strictest setting flagged 11 rewrites (5%), down from 26 for the first release. On 60 of those drafts, the popular blader/humanizer prompt applied by Claude Sonnet was flagged 60 times out of 60; this model, 4. These are the author's own measurements on one detector on one date, and the card says as much.
The more useful number is fact fidelity. A strict LLM judge (GLM-5.3) found no factual problem in 376 of 420 English rewrites from the recommended Q8_0 file, which means about one in ten has an error. The author says more than nine in ten of the flagged fixes are a single word or phrase, and gives a real example: "the remaining 37 complaints" became "the other 37% of complaints". Chinese is weaker, with 149 of 204 clean. Formatting slips too: 28 of 420 outputs dropped a list, heading or break.
The engineering is better than most hobby releases. The 2-bit file was rebuilt with quantisation-aware training so it matches the full model's next token 87% of the time, against 70% for a standard 2-bit build of the same size, and runs on an 8 GB machine. The card documents the exact prompt byte for byte, the sampling that must be switched off, and a self-test. On an M5 Max a 100-word email takes about 3.6 seconds.
It is still a tool whose main market includes students and anyone required to disclose AI help. The author asks users to follow school and employer rules, but nothing in the model enforces that. Used on your own drafts, to make an AI-assisted email sound like you, it is a well-made local alternative to pasting text into a web humaniser.
Where it earns its place
A sales manager at a 20-person agency who drafts client emails with an AI assistant
Make AI-drafted emails sound less templated without leaking client details
The Mac or Windows app downloads one file once and then works offline; a 100-word email takes about 3.6 seconds on an M5 Max. Marking a passage as a fact keeps it word for word, but numbers and names still need a read.
A developer running a documentation pipeline
Rewrite generated release notes in a plainer voice in batch
Serve the Q8_0 GGUF with llama-server and call /completion with the documented prompt, or run hz on .md files; it keeps headings, code and tables and flags any section where a number went missing. Split files to stay within 8,192 tokens.
Before you commit
- Setup
- Install the desktop app, which suggests one of five file sizes for your memory and downloads it, or fetch a GGUF with hf download and run llama-server. The prompt format must be reproduced exactly and top-k and min-p switched off; the card gives a hash to check it.
- Integrations
- llama.cpp, LM Studio, Ollama (with the supplied Modelfile), MLX, transformers and vLLM; an AGENTS.md file tells coding agents how to install and self-test it. A community draft model speeds MLX decoding about 1.5x.
- Data and privacy
- Runs entirely locally; nothing is sent anywhere. Apache 2.0, inherited from Gemma 4 12B; the author is not affiliated with Google.
- Cost at scale
- Free. Q8_0 is 12.7 GB with a 13.7 GB peak on a Mac, Q4_K_M is 7.6 GB, and the 2-bit file is 3.9 GB with a 6.2 GB peak; speed on an M5 Max is about 36 to 38 tokens a second.
- Limits
- 8,192 tokens per request, so long documents must be split. The card lists wrong-language replies on short jargon-heavy English drafts, added slang or profanity in casual genres, and templated social posts as the genre detectors still catch (3 of 16).
What’s good
- 376 of 420 English rewrites with no factual error (self-reported)
- Runs offline, down to 8 GB of memory with the 2-bit file
- Training data, prompts and evaluation outputs all published
- Apache 2.0; free Mac and Windows app
What’s not
- About one rewrite in ten still alters a fact
- Detector results are one vendor on one date
- Chinese and formatting are less reliable
- Sometimes answers in Chinese or adds slang unprompted
Best for
People who draft with AI and want the result to sound like their own writing, on their own machine, and who will proofread numbers and names before sending.
humanizer (12B) alternatives
humanizer (12B): common questions
- Is humanizer (12B) worth it?
- Unusually well documented and better at keeping facts than a prompt, but one English rewrite in ten still changes something, and its headline score rests on a detector that may change tomorrow. We rate it 3.7 out of 5.
- How much does humanizer (12B) cost?
- Free under Apache 2.0: weights, GGUF files, desktop app and CLI. No hosted service.
- What are the downsides of humanizer (12B)?
- About one rewrite in ten still alters a fact; detector results are one vendor on one date; chinese and formatting are less reliable; sometimes answers in Chinese or adds slang unprompted.
- Who is humanizer (12B) best for?
- People who draft with AI and want the result to sound like their own writing, on their own machine, and who will proofread numbers and names before sending.
- How much does humanizer (12B) cost for a team or in production?
- Free. Q8_0 is 12.7 GB with a 13.7 GB peak on a Mac, Q4_K_M is 7.6 GB, and the 2-bit file is 3.9 GB with a 6.2 GB peak; speed on an M5 Max is about 36 to 38 tokens a second.
- Is humanizer (12B) safe for company data?
- Runs entirely locally; nothing is sent anywhere. Apache 2.0, inherited from Gemma 4 12B; the author is not affiliated with Google.
- What are the best humanizer (12B) alternatives?
- The closest alternatives we have reviewed are Gemma 4.
Answers checked Oct 31, 2026.