Humanizer Rewrites AI Text Locally So Detectors Flag Just 11 of 210
A locally-run 12B model fine-tuned from Gemma-4 rewrites AI drafts into natural human prose while preserving every number, name, and quote.
- humanizer is a 12B Gemma-4 fine-tune that rewrites AI text to sound human, in English and Chinese
- Apache 2.0 weights, GGUF/safetensors/MLX, runs locally via Hugging Face or desktop app
- 95% of rewrites judged human by Originality.ai at strictest setting, vs 88% for previous release
- 376 of 420 English rewrites had no factual problem per strict LLM judge on Q8_0 file
- Trained with SFT, DPO and three rounds of GRPO; no AI detector used anywhere in training
- About 36-38 tokens/sec on M5 Max via llama.cpp Metal; 8192 token context window
Humanizer targets LLM cadence with a local 12B model
Humanizer, a new open-weights model with downloadable parameters, has climbed Hugging Face’s trending list. The 12-billion-parameter model rewrites AI-generated emails, essays, reports, forum posts, and other drafts in English or Chinese. Its training objective favors natural phrasing while preserving every number, unit, date, name, and quotation without adding new information.
Downloadable weights let developers process sensitive drafts locally and integrate rewriting into existing pipelines. The training loop used no AI detector as a reward signal. Originality.ai appears only in the published evaluation, while training rewards focus on factual fidelity and avoiding close copies of reference text.
Weights for four hardware tiers
The release fine-tunes Google’s Gemma 4 12B and ships under Apache 2.0. GGUF quantizations support llama.cpp, safetensors support Transformers and vLLM, and an MLX path targets Apple Silicon. Quantization reduces numerical precision to lower memory and storage requirements.
| Variant | Approximate size | Project guidance |
|---|---|---|
humanizer-12b-Q8_0.gguf |
12.7 GB | Machines with at least 32 GB of memory |
humanizer-12b-Q6_K.gguf |
10 GB | Machines with 16 GB of memory |
humanizer-12b-Q4_K_M.gguf |
7.6 GB | Systems with tighter memory or storage limits |
lite/ |
Earlier E4B release | Machines with 8 GB of memory |
The project also provides a desktop application for macOS and Windows. Its hz command-line tool processes .md, .txt, and .docx files, preserving headings, code, tables, and links. For long documents, it splits text at paragraph boundaries and resamples chunks when a number disappears.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
