Humanizer 12B Rewrites AI Text to Bypass Detectors

A new open-weights model called Humanizer has emerged on Hugging Face, designed to rewrite AI-generated drafts so they read as if they were written by humans. Built as a 12-billion-parameter fine-tune of Google's Gemma 4, the model operates locally and supports both English and Chinese. In evaluations using the strict settings of Originality.ai, the model achieved a 95 percent human-score rating, up from 88 percent in its previous release. Out of 210 tested rewrites, AI detectors flagged only 11.

Unlike other rewriting tools, Humanizer was trained without using an AI detector as a reward signal. Instead, its training pipeline relied on supervised fine-tuning, direct preference optimization, and three rounds of group relative policy optimization. The training rewards prioritized factual accuracy and avoided close copying of the source text. As a result, a strict LLM evaluation of the Q8_0 file format showed that 376 out of 420 English rewrites contained no factual errors, preserving critical data like names, dates, and numbers.

For practitioners, the model offers highly flexible deployment options under an Apache 2.0 license. It supports GGUF quantizations for llama.cpp, safetensors for Transformers and vLLM, and MLX for Apple Silicon. On an M5 Max chip running llama.cpp Metal, the model achieves processing speeds of 36 to 38 tokens per second within an 8,192-token context window.

The release also includes a desktop application for macOS and Windows, alongside a command-line tool called hz. This utility allows developers to process markdown, text, and Word documents directly while preserving formatting elements like tables, code blocks, and links. For longer documents, the system automatically segments text at paragraph boundaries to maintain coherence and resamples chunks if any numerical data is lost during the rewrite.