---
type: "post"
id: "humanizer-a-local-12b-model-that-rewrites-ai-b81290"
slug: "humanizer-a-local-12b-model-that-rewrites-ai-b81290"
title: "Humanizer: A local 12B model that rewrites AI-written drafts so they sound more human without changing the facts."
description: "Humanizer’s model card describes a 12-billion-parameter rewriting model based on Google Gemma 4 12B, with English and Chinese listed and an Apache-2.0 license."
retrieval_nugget: "Humanizer’s model card describes a 12-billion-parameter rewriting model based on Google Gemma 4 12B, with English and Chinese listed and an Apache-2.0 license."
published_at: "2026-10-16"
updated_at: "2026-10-16"
record_date: "2026-10-16"
date_kind: "published_at"
topics: ["google"]
source_urls: ["https://huggingface.co/jialinyyzz/humanizer"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:humanizer-a-local-12b-model-that-rewrites-ai-b81290"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:96c7838371569ba2e1000cd3d1888767038bf419d3b2be4e16a4414c6f720df4","reviewed_at":"2026-10-16T20:11:23.635Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"A narrowly prompted rewriting model can make local style editing accessible, but its promise of factual preservation remains an instruction to test, not a verified property.","observed_facts":[{"text":"The Hugging Face model card identifies Humanizer as a 12-billion-parameter model based on Google Gemma 4 12B and lists English and Chinese.","source_urls":["https://huggingface.co/jialinyyzz/humanizer"]},{"text":"The card lists an Apache-2.0 license, Safetensors and GGUF weights, quantized files, and usage paths through Transformers and llama.cpp.","source_urls":["https://huggingface.co/jialinyyzz/humanizer"]},{"text":"The prompt instructs the model to preserve facts, numbers, units, dates, names, and quotations while rewriting.","source_urls":["https://huggingface.co/jialinyyzz/humanizer"]}],"mechanism":"The prompt directs stylistic restructuring while GGUF packaging and llama.cpp enable local inference; the preservation constraint is expressed in the prompt rather than verified by supplied evaluations.","why_now":"The release packages a specific editing task with weights and documented local serving paths, giving teams a practical option to evaluate in their existing workflows.","implications":["Editorial teams can trial a bounded rewrite step locally without sending drafts to a hosted service.","Fact-checking or source comparison remains necessary because the supplied evidence does not establish preservation reliability."],"evidence_boundary":"The supplied model card documents packaging, prompt intent, and usage instructions, but provides no benchmark or validation showing that rewrites reliably preserve facts.","watch_conditions":["Look for a reproducible evaluation reporting factual retention across diverse drafts and languages."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied for this particular model release; it merits a distinct record about locally packaged rewriting and the gap between prompt-level constraints and demonstrated reliability."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/humanizer-a-local-12b-model-that-rewrites-ai-b81290/","markdown":"https://newruntime.com/posts/humanizer-a-local-12b-model-that-rewrites-ai-b81290.md","json":"https://newruntime.com/posts/humanizer-a-local-12b-model-that-rewrites-ai-b81290.json"}
---

# Humanizer: A local 12B model that rewrites AI-written drafts so they sound more human without changing the facts.

## Retrieval answer

Humanizer’s model card describes a 12-billion-parameter rewriting model based on Google Gemma 4 12B, with English and Chinese listed and an Apache-2.0 license.

Humanizer’s model card describes a 12-billion-parameter rewriting model based on Google Gemma 4 12B, with English and Chinese listed and an Apache-2.0 license. It offers Safetensors and GGUF weights, including quantized files, and documents local use through Transformers and llama.cpp. This is a tangible packaging choice: rewriting can be tested as a local component rather than only as a hosted feature.

The thesis: A narrowly prompted rewriting model can make local style editing accessible, but its promise of factual preservation remains an instruction to test, not a verified property. The mechanism: The prompt directs stylistic restructuring while GGUF packaging and llama.cpp enable local inference; the preservation constraint is expressed in the prompt rather than verified by supplied evaluations. Why now: The release packages a specific editing task with weights and documented local serving paths, giving teams a practical option to evaluate in their existing workflows.

Editorial teams can trial a bounded rewrite step locally without sending drafts to a hosted service. The evidence boundary: The supplied model card documents packaging, prompt intent, and usage instructions, but provides no benchmark or validation showing that rewrites reliably preserve facts. The first implication: Editorial teams can trial a bounded rewrite step locally without sending drafts to a hosted service. The first watch condition: Look for a reproducible evaluation reporting factual retention across diverse drafts and languages.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)