Betty is a fully local, offline copy-editing app for book-length manuscripts that runs Qwen3.5 4B, combining overlapping-chunk processing, LanguageTools/HunSpell spell checkers, and a writer-in-the-loop accept/dismiss interface, with results benchmarked against planted errors across five languages.
Hey y'all! I have a pet project that has grown out of proportions. Long story short: I'm a data scientist who writes fantasy books and self-publish them. I think it's a genuine waste of human life to check for spelling errors so I figured AI could help. Turns out, it is not so simple to get an AI to properly fix a 120k words manuscript ;)... That's why I created Betty! It runs Qwen3.5 4B as an offline copy editor for whole novels — and this is some of what I learned fighting tens-of-thousands of words through a 4B model, insisting that (most) users can use it fully for free and fully offline. Because, let's be honest: authors rightfully distrust and generally hate AI companies. First challenge: Chunking the text. Authors already do this, in the darkest hours of the night, copy-pasting snippets into chatGPT for some shameful feedback. Problem with that approach: super inefficient, both for the author and the environment. And the context is missing at the edges of each chunk. I fixed it by ensuring chunks to overlap. Second challenge: AI misses genuine errors. So, I added two conventional spell controllers -- LanguageTools and HunSpell. This already surfaces all spelling errors, letting the AI focus on suggested fixes and on all the "non-error" errors, such as "There" vs "Their". For these, the AI searches, while a Python script surfaces all the common culprits for the model to pay special attention. Third challenge: Error rate. First off, Betty doesn't capture everything. Second off, it sometimes introduces its own mistakes. I fix it by putting the writer-in-the-loop, and there's a super smooth interface now for the author to accept and dismiss suggested edits (tinder-style with left and right swipes ; ) ). I'd be super grateful for any advice, feedback, and thoughts you might have on this project. I currently have it up-and-running with about 30 users and getting some user feedback. Northing technical though, so this is what I'd love to have more of. Full thing is source-available on GitHub, and can be found for download and lots more information at www.bethaniel.eu
The author is highly impressed with Qwen 3.8 27B's German translation capabilities, noting its superior vocabulary, sentence structure, and punctuation compared to other frontier models like GPT-5.6 and Fable 5.
Fixed three bugs in a qMLX fork for running Qwen3.5-122B on Mac Studio, reducing prefill time from minutes to sub-seconds for long-context inference; open-sourced the fork and benchmark script.
A small open model, qwen3-4b-deslop, is a Qwen3-4B-Instruct fine-tune trained on a MacBook Air that removes common AI writing tics (em dashes, 'not X, but Y', lists of three), cutting the slop score from 121 to 12.5 per 1000 words while keeping factual accuracy, and it runs locally via Ollama.
Nativ is an open-source macOS app that runs Qwen 3.5 9B and other open models locally on Apple Silicon, offering customizable system prompts, telemetry, and integrations with coding agents, with no cloud or subscription required.