Skip to content
← Models
Open weightsjialinyyzz/10 trustMetrics-checked

Humanizer

A 12B open-weights model that rewrites AI-written drafts so they read like a person wrote them, keeping every number, name, and quote intact.

Visit official page →

About

Humanizer is a 12B-parameter finetune of google/gemma-4-12B trained specifically to rewrite AI-written drafts — emails, essays, reports, forum posts in English and Chinese — so they read like a person wrote them. Its training objective is to preserve every number, unit, date, name, and quote while adding nothing, and the author publishes held-out before/after examples with full eval outputs on GitHub. At its strictest setting, Originality.ai judged 95% of its English rewrites human (11 of 210 flagged), though those results are provider-reported and the author advises re-reading numbers before sending. It ships as GGUF and safetensors with a free desktop app for Mac and Windows that runs fully offline, and also works with stock llama.cpp. Released 2026-09-11 under Apache 2.0, it has drawn about 85.6k downloads and 1.8k likes on Hugging Face.

Specs

Context window

Not specified on the model card

Pricing

Free — open weights, run locally at your own compute cost

Free quota

Weights free under Apache 2.0; no managed API; free desktop app for Mac/Windows

API access

No managed API — download GGUF/safetensors and run via llama.cpp, or use the free offline desktop app

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Originality.ai (strictest setting): 95% of English rewrites judged human, 11 of 210 flagged (previous release 26) — provider-reported (own eval set)
  • Held-out eval: fact judge flagged 44 of 420 English rewrites — provider-reported

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

writingautomation

Trust score — 5/10

/10 trustMetrics-checked

Scored 7/10 (METRICS tier, metrics-v1). Evidence: 1,829 likes and ~85.6k downloads; Apache 2.0; full eval outputs published on GitHub with unedited best picks; desktop app and GGUF builds make local use trivial. Docked for: (1) solo individual author, so maintenance is a single-person dependency; (2) headline humanization scores are provider-run with no independent replication seen. License: Apache-2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · HF: 1,829 likes, ~85.6k downloads, created 2026-09-11
  • · manual review: base model, license, eval method verified; headline humanization scores are provider-run
  • · docked: solo individual author is a single-person maintenance dependency

Sources