42ailab/AIWriter-V1-4B-Check-Preview

AIWriter Chinese Proofreading Model · V1 Preview

AIWriter-V1-4B-Check-Preview

English | 简体中文 · Also on ModelScope · 42ailab.com

⚠️ Preview release. V1 is not finalized. The stable release will be published at 42ailab/AIWriter-V1-4B-Check after the AIWriter launch event. This repository will be kept, not deleted (for reproducibility) and marked as superseded at that time. "V1" denotes the model generation, not a finalized version.

---

1. Introduction

AIWriter Chinese Proofreading Model is trained in-house by 42ailab for serious long-form Chinese writing: typos, word misuse, grammar, and punctuation.

What sets it apart from general-purpose LLMs doing proofreading is what it optimizes for:

General models optimize for finding everything. This model optimizes for holding back.

We measured three leading flagship models on sentences that contain no error at all — 43.4%–50.4% of them got edited anyway. (A fourth, kimi-k3, scored highest on the main metric but its false-edit rate has not been measured yet; it is listed in the table below with that column blank rather than silently dropped.)

For an author this is worse than a missed typo. A missed typo still gets caught by the editor; a quietly "improved" sentence stays in the book — it reads fine, it just isn't the sentence you wrote anymore.

Key Features

• Low false-edit rate: 19.92% on error-free sentences — less than half that of the three flagships we measured (43.4%–50.4%);
• Guaranteed by training, not by prompting: "return the sentence unchanged when there is no error" is a behaviour learned from the training distribution, not an instruction in a system prompt;
• Purpose-built benchmark: accompanied by AIWriter CheckBench, which treats the false-edit rate as a first-class metric alongside the main score;
• Fully local: 4B parameters, runs on an ordinary laptop, works offline — your manuscript never leaves your machine;
• Open weights: free for academic research and personal use.

2. Model Summary

Base model	Qwen3-4B (Apache-2.0, Qwen Team)
Parameters	4B
Context length	4,096
Recommended output	192 tokens
Language	Simplified Chinese
Released formats	GGUF Q8_0 (4.28 GB) · Q5_K_M (2.89 GB)
License	AIWriter Model License 1.0 (free for academic & personal use; commercial use requires a licence)

3. Evaluation

Results on AIWriter CheckBench · V1 Preview, a Chinese proofreading benchmark built in-house by 42ailab. Target measures whether edits are correct; False-edit measures whether error-free sentences were left alone:

Model	Target	False-edit ↓
kimi-k3	63.17	—
DeepSeek-V4.1-Flash	59.34	45.3%
GLM-5.2	56.99	50.4%
qwen3.8-flash	55.83	43.4%
This model (4B)	50.53	19.92%

Read the two columns separately. The flagships lead by 5–13 points on Target, while their false-edit rate is 2.2–2.5× higher than this model's.

Scope: figures are from the public development set; baselines were measured in 2026-09 on the same set, with the same scorer and the same decoding settings. A held-out test set is reserved for the closed evaluation of the stable release, following standard benchmark practice.

The full evaluation design, training recipe and ablations will appear in the technical report and the accompanying papers.

4. Download & Deployment

This model is packaged for 42model, our local inference engine — the recommended way to get it:

42model desktop

1. Open Model Library → Writing and download AIWriter Chinese Proofreading Model · V1 Preview;
2. Pick a build:
   • Q8_0 (4.28 GB) — recommended, closest to the original weights;
   • Q5_K_M (2.89 GB) — 1.4 GB lighter, 0.7 points lower.

Downloading is all you need to do here — AIWriter picks the model up on its own (next section).

Inside AIWriter

Once installed, open AIWriter — the desktop app discovers the local engine automatically; no endpoint or key to configure. Click Proofread in the editor. Everything runs on your machine; your manuscript never leaves it.

Files

File	Build	Size
AIWriter-V1-4B-Check-Q8_0.gguf	quality	4.28 GB
AIWriter-V1-4B-Check-Q5_K_M.gguf	lightweight	2.89 GB

SHA-256 checksums are listed on the Files page.

5. Limitations

• It is a sentence-level proofreader; long documents must be segmented and reassembled by the application;
• It fixes objective errors only — no stylistic rewriting or polishing;
• Training data is predominantly Simplified Chinese formal prose; internet slang, dialects and classical Chinese have not been systematically evaluated.

6. License & Attribution

• This model is released under the AIWriter Model License 1.0: free for academic research, teaching, evaluation and personal use (including fine-tuning and derivatives, with attribution); commercial use requires prior written permission (contact@42ailab.com).
• Intellectual property: the training methodology, data construction and evaluation design remain the property of 42ailab. This licence grants no patent rights.
• Base model: Qwen3-4B, © Qwen Team, Apache-2.0. Rights in the base model are unaffected by this licence — anyone may obtain and use Qwen3-4B itself under Apache-2.0; this licence applies only to the part we trained on top of it.
• Training data comes from 42ailab's own corpus, synthetic data and public datasets.

7. Citation

The technical report and accompanying papers are in preparation; DOI and arXiv links will be added here on publication. In the meantime:

@misc{aiwriter2026check,
  title        = {AIWriter Chinese Proofreading Model V1 (AIWriter-V1-4B-Check)},
  author       = {Yang, Zhiping and {42ailab}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/42ailab/AIWriter-V1-4B-Check-Preview}},
  organization = {42ailab},
  note         = {Preview weights; technical report forthcoming, DOI and arXiv to be added here.
                  Base model Qwen3-4B (Apache-2.0). Contact: contact@42ailab.com}
}

Contact: contact@42ailab.com

About Us

42ailab — an AI research lab exploring the boundaries of intelligence. Grounded in cognitive science, we work towards a deep integration of AI and human intelligence — truly understanding and augmenting intelligence, carbon-based and silicon-based alike.

42model — a high-performance local AI inference engine from 42ailab. Translation, transcription, recognition, chat and coding run free and private on your own machine; you can also fine-tune your own model with cloud compute and bring the weights back home.