A small open model trained to drop AI writing habits like the em dash

Erik Markus Kannike has released qwen3-4b-deslop, a version of the small open Qwen3-4B model fine-tuned on a MacBook Air to avoid common AI writing habits: em dashes, "it's not X, it's Y", lists of three and words like "delve". It needs no special prompt and runs on a laptop with Ollama as a 2.5 GB download. On 40 test prompts it had not seen, his own habit score fell from 121 to 12.5 per 1,000 words, em dashes went to zero, and it still got 30 of 30 factual questions right. He is open about the limits: the voice is flat, it still invents names and places, answers run long, the score is his own rule-based detector with no blind test by readers yet, and he has only tested short English prose. On Hacker News (145 points, 107 comments), 5 of the 33 top-level commenters say the output still reads as AI-written or is worse in places (one says it still scores 100% AI on the Pangram detector), and one tester found the 4-bit build starts looping after about 16,000 generated tokens. Three praise the result or how little hardware it needed. Three object to the goal: one wants AI text to stay easy to spot, and two defend em dashes and lists of three. Seven ask questions, including three requests for the detector or style guide, five comment on the page's live cursors and design, and nine add side ideas, including an earlier "antislop" Gemma model. No consensus, and the thread leans skeptical. Sorted take: A useful experiment rather than a writing tool; for work drafts, a clear house style guide and a human edit still do more than a small model trained to avoid a list of tics.