Posted as a Show HN on October 19, 2026, qwen3-4b-deslop is an Apache-2.0 fine-tune of Qwen3-4B-Instruct whose author trained it, on a MacBook Air, to drop the stereotyped habits that mark a draft as machine-written — with the results measured by a home-grown "slop score."
2026-10-19 · by Moe Ameen
On October 19, 2026, a developer named Erik Markus Kannike posted a Show HN for qwen3-4b-deslop — a fine-tune of Alibaba's small open model Qwen3-4B-Instruct-2507 whose entire purpose is to write prose without the giveaways that read as AI-generated. The model card names the targets directly: em dashes, "not X, but Y" constructions, and lists of three. It is released on Hugging Face under the base model's Apache-2.0 license as full weights (`erikkannike/qwen3-4b-deslop-v4`), a 2.5 GB Q4_K_M GGUF for Ollama and llama.cpp, and a 126 MB LoRA adapter.
The method is distillation, and the author is candid about running the whole thing on an M2 MacBook Air at one to three hours per round. A teacher model answers 288 training prompts, answers that still show the tics or miss the task are filtered out, and a rank-16 LoRA adapter learns to produce the survivors from the bare prompt. It took three rounds: an intermediate version scored better on style but started writing opinion pieces with no opinion, inventing studies, and joining clauses with ", which" — so round three added a second filter checking that an answer actually did what the prompt asked. The release also ships an unusual 175-byte "neuron edit" that flips the sign of four MLP neurons, which the author says removes roughly 70% of the "not X, but Y" pattern on its own.
The headline numbers are measured by the author's own regex-based "slop score" on 40 held-out prompts, and he is clear that no blind human test has validated it. On that metric the score fell from 121 to 12.5 per 1,000 words (human writing, he notes, usually sits under 15); em dashes went from 20.5 to zero, "not X, but Y" from 7.8 to 0.8, and lists of three from 17.3 to 1.1. Factual questions stayed correct (30 of 30), task fit edged up to about 80%, and 68% of answers landed in his "human range" versus 0% for the original. The honest caveats are in the card too: the voice turns flat and diary-like, it invents names and places, answers run long, and only short English prose in four genres was tested — code, other languages, and tool use were not.
The useful takeaway here is not "go self-host a 4B model" — it is that fighting AI-tell has become a defined, repeatable discipline, and the place it pays off is wherever you generate copy at volume. [Kompozy](/) treats that as a default, not a bolt-on: every text output it produces — Text Posts, captions, a [Blog Article](/ai-tools/koala-ai), an Email Newsletter, the copy on a carousel — is governed by a [Persona Brief](/glossary/persona-brief) plus banned-word filters, so the same tics this project trains out (robotic phrasing, filler triads, hedged non-statements) are screened at the point of generation and the voice stays consistent across formats instead of being de-slopped by hand one post at a time.
That consistency is the part a single local model can't give a creator. qwen3-4b-deslop cleans one block of short English prose in a chat window; a content week is a captioned [Persona Shorts](/glossary/persona-shorts) script, a brand-exact [Carousel](/glossary/hyperframes), quote graphics, native posts written per platform, a blog, and a newsletter — all of which need to sound like one human. Kompozy generates that whole set from one seed and then schedules and publishes it across the eight social platforms plus blog and email behind a per-post review, so your on-brand, un-robotic voice ships everywhere at once. (Kompozy's own copy runs on Claude and OpenAI, so this is about the shared goal — human-sounding output — not a plug-in for this specific model.) If you like drafting locally, keep doing it and paste the result into Kompozy as a seed; the engine turns a de-slopped paragraph into a published, multi-format campaign.
It is a fine-tune of Alibaba's open Qwen3-4B-Instruct-2507 model, released by developer Erik Markus Kannike in October 2026, trained specifically to write without common AI-assistant tics such as em dashes, "not X, but Y" phrasing, and lists of three. It ships on Hugging Face under the Apache-2.0 license as full weights, a 2.5 GB GGUF for Ollama, and a LoRA adapter.
On the author's own regex-based "slop score" the improvement is large — 121 down to 12.5 per 1,000 words, with em dashes eliminated — and 68% of its answers fell in his "human range." But he is clear that no blind human test has validated the metric, the base is only a 4B model, and it can go flat, invent names and places, and run long. Treat it as a capable drafting aid rather than a finished-copy generator.
By distillation across three rounds on an M2 MacBook Air: a teacher model answered 288 prompts, answers that still showed the tics or missed the task were filtered out, and a rank-16 LoRA adapter learned to reproduce the clean answers. A later round added a check that the answer did what the prompt asked, and the release includes a small "neuron edit" that flips four MLP neurons to further suppress the "not X, but Y" habit.
No. Self-hosting a 4B model is one route and a good tinkering project. For a creator shipping real content, the practical move is to make human-sounding output a default in your generation pipeline. Kompozy, for example, governs every text format with a Persona Brief and banned-word filters so the same tics are screened at generation and the voice holds across captions, blogs, carousels, and newsletters — then publishes the result across every platform.