Humanizer ships app 0.3.3 for its local 12B AI-draft rewriter

What shipped on 10 October?

The sgaofen/humanizer-local-model repository published GitHub release app-v0.3.3 on 10 October 2026 with Humanizer-0.3.3-macos-arm64.dmg and Windows setup/portable builds. The release notes focus on post-rewrite editing: the draft no longer becomes read-only after a rewrite, Create fact passages are highlighted in both panes, and a one-line hint explains how to mark text that must survive unchanged.

That app release sits on a busier week for the same project. App 0.3.0 on 6 October added Q3 and 2-bit sizes for roughly 8 GB machines; 0.3.1 fixed occasional English-to-Chinese rewrites; 0.3.2 on 7 October added .docx/PDF import, Create fact, and update checks. Readers should treat 10 October as the latest app milestone, not as the first day the 12B weights existed—the Hugging Face model card shows an earlier September create time with ongoing file updates.

For adjacent local tooling already covered here, compare open weights versus open source when you evaluate Apache-2.0 GGUF drops, and AI skills for teams if you plan to wrap the rewriter behind a shared agent skill.

What does the 12B model actually do?

Humanizer is a specialized rewriter, not a general assistant. The instruction asks the model to rewrite so the text “reads like a person wrote it,” to vary sentence length, cut hedging, prefer concrete words, and keep every fact, number, unit, date, name, and quotation unchanged. The prompt format is stored in prompt_format.json and must be reproduced byte-for-byte for chat front ends.

Weights live at jialinyyzz/humanizer under Apache 2.0, fine-tuned from google/gemma-4-12B according to the model card tags. Distributed files include bf16 safetensors, a 23.8 GB bf16 GGUF, and Q8_0 / Q6_K / Q4_K_M / Q3-QAT / IQ2_XS-QAT GGUF builds. The README’s size table reports peak Mac memory while rewriting and top-1 agreement versus bf16; those hardware numbers are the authors’ measurements on their machines.

Create fact is an app-layer guarantee attempt: marked passages are swapped for placeholders the model is expected to copy, then restored. The 0.3.2 notes claim 130 marked passages across 64 rewrites returned word-for-word on the first try in their test—again, author-reported, not reproduced here.

How should readers treat the evaluation numbers?

The README’s English results card states that on 2 October 2026, with bf16 weights, 199 of 210 English rewrites (95%) were judged human-written by Originality.ai at its strictest setting, improving on 184/210 for a previous release in their comparison. Separately, a strict LLM fact judge flagged problems in 44 of 420 English rewrites on the Q8_0 file users download, with most issues described as single-word or single-phrase slips.

Those figures are useful as the vendor’s published scoreboard and methodology pointers (eval/ on GitHub), not as transferable proof that your detector, genre, or language mix will match. Detectors change; the project itself says to read numbers, dates, and names before sending. Chinese fact-judge results in the same README remain weaker than English in the authors’ own table.

Quantization-aware training for the smaller GGUF files is presented as improving top-1 agreement with bf16 versus a naïve llama-quantize pass—at 2-bit, 87% versus 70% in their chart. That is an author benchmark about next-token agreement, not a claim that 2-bit files are as fact-safe as Q8_0.

How do you run it without trusting the app binary?

The project documents three paths: the Electron-style local app that downloads a GGUF once; a one-command llama-server pointing at the Hugging Face repo; and hz, a Python CLI that rewrites Markdown or .docx piece by piece while checking for lost numbers. AGENTS.md is aimed at coding agents that should pin files, prompt bytes, and a self-test.

Security-conscious teams should still treat unsigned desktop apps cautiously—the INSTALL guide admits the app is not code-signed yet—and prefer pinning hashes from USAGE.md. Pair local rewrite with sandboxing patterns such as GitHub Copilot local sandboxing if the same machine also runs coding agents with tool access.

Because the model is text-completion only, glue code that silently wraps chat templates can degrade quality. If an integration “doesn’t sound right,” verify the prompt against prompt_format.json before blaming the weights.

1. Install app 0.3.3 or pin a GGUF SHA from USAGE.md before sharing with a wider team.
2. Run Create fact on a draft that contains invoice numbers and proper names; confirm they survive.
3. Record which detector, if any, you care about—and re-measure after any model file update.

What are the practical limits?

Style transfer is not authorship verification. A document that passes one detector can still be false, plagiarized, or inappropriate for regulated disclosure. Humanizer’s own framing is editorial rewriting with a fact-preservation instruction, not a legal attestation pipeline.

Long documents may drift: the app warns that long imports need review. Smaller quants trade fact slips for RAM. The model is bilingual for English and Chinese rewrite tasks in the authors’ materials; other languages are outside the claims we inspected.

Finally, “humanizer” tools sit in a contested category. Using them to evade classroom or publisher AI policies may violate local rules even when the software license permits the download. That is a policy constraint on the user, not a defect in Apache 2.0.