ericthewiz/humanizer

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The ericthewiz/humanizer is a 12 billion parameter model, fine-tuned from Google's Gemma-4-12B, specifically designed to rewrite AI-generated text (emails, essays, reports, forum posts) to sound more human. It maintains factual accuracy, preserving numbers, units, dates, names, and quotes, and supports both English and Chinese. This model excels at making AI drafts indistinguishable from human writing, achieving 95% human detection rates on Originality.ai, and is optimized for local deployment with a 32768 token context length.

Loading preview...

humanizer: Making AI Text Sound Human

The ericthewiz/humanizer is a 12 billion parameter model, fine-tuned from Google's Gemma-4-12B, engineered to transform AI-written drafts into natural, human-sounding prose. Unlike general-purpose LLMs, humanizer is specialized in rewriting text while strictly preserving all factual details like numbers, dates, and names, ensuring no information is lost or altered. It supports both English and Chinese.

Key Capabilities

  • Human-like Rewriting: Achieves a 95% "judged human" rate by Originality.ai (strictest setting), significantly outperforming previous versions and other tools.
  • Factual Fidelity: Designed to keep every number, unit, date, name, and quote unchanged, with 376 out of 420 English rewrites having no factual problems.
  • Multilingual Support: Effectively rewrites drafts in both English and Chinese.
  • Optimized Quantization: Features advanced GGUF quantization (2-bit, Q3, Q4_K_M) that maintains high agreement with the full bf16 model at significantly smaller file sizes, enabling efficient local deployment.
  • Local Operation: Runs efficiently on consumer hardware, with an app available for Mac and Windows, and support for llama.cpp, MLX, transformers, and vLLM.

Good For

  • Developers integrating human-like text generation into applications or agents.
  • Content creators needing to refine AI-generated drafts for a more authentic voice.
  • Anyone looking to make AI-written emails, reports, essays, or social media posts sound genuinely human without losing critical information.
  • Users with 8GB+ memory who want to run a specialized text rewriting model locally and offline.

Finetunes
1 models