An Apache-2.0 fine-tune of Qwen3-4B-Instruct trained to write without the usual AI tics — em dashes, "not X, but Y", and lists of three — runnable locally as a 2.5 GB file in Ollama.
Last verified · 2026-10-19 · by Moe Ameen
qwen3-4b-deslop is a small, open writing model: a fine-tune of Alibaba's Qwen3-4B-Instruct-2507 built by developer Erik Markus Kannike and posted as a Show HN on October 19, 2026. Its single goal is prose that does not read as machine-written. The model card names the targets outright — em dashes, "not X, but Y" constructions, and lists of three — and the fine-tune drives them down without a system prompt or any special scaffolding.
It was made by distillation, and the author ran the whole thing on an M2 MacBook Air. A teacher model answers 288 prompts, answers that still show the tics or miss the task are filtered out, and a rank-16 LoRA adapter learns to reproduce the clean ones; it took three rounds, with a task-fit filter added in round three to stop an earlier version writing contentless opinion pieces. The release also ships a tiny "neuron edit" that flips four MLP neurons to further suppress the "not X, but Y" habit. On the author's own regex "slop score" over 40 held-out prompts, the number fell from 121 to 12.5 per 1,000 words, em dashes to zero, and factual answers stayed correct (30/30) — though he is clear the metric is self-reported with no blind human test, and that the model can go flat, invent names and places, and overrun the requested length.
It is free under the base model's Apache-2.0 license and comes three ways on Hugging Face: full bf16 weights for transformers, a 2.5 GB Q4_K_M GGUF to run in Ollama or llama.cpp, and a 126 MB LoRA adapter. It is a text model only — no images, video, or publishing. Because it is a fast-moving hobby release, confirm the current version, metrics, and files on its Hugging Face page before relying on specifics.
The honest shape of this tool is narrow and useful: it hands you a cleaner block of text in a chat window. What it never does is turn that paragraph into the thing your audience actually sees — a captioned video, a branded carousel, a scheduled post. That is the handoff [Kompozy](/) is built for, and the two fit together cleanly because they share a goal: output that doesn't sound robotic.
Here is the concrete loop. Draft locally with qwen3-4b-deslop — a hook, an outline, a rough script — then paste the de-slopped text into Kompozy's Quick Ingest as a seed. Kompozy fans that one seed into the formats the model can't make: a captioned [Persona Shorts](/glossary/persona-shorts) avatar video that narrates the script, a brand-exact [Carousel](/glossary/hyperframes), Quote Graphics of the sharpest line, native Text Posts written per platform, a Blog Article, and an Email Newsletter. Crucially, every one of those is held to a single voice by the [Persona Brief](/glossary/persona-brief) plus banned-word filters — so the un-robotic tone you got in one paragraph is enforced across the whole set instead of degrading the moment you leave the box where you tested it. Then [Autopilot](/glossary/autopilot) schedules and publishes the lot across the eight social platforms plus blog and email behind a per-post review. (Kompozy's own generation runs on Claude and OpenAI, so this is an upstream drafting habit you bolt on, not a model swap.) The model gives you a clean sentence; Kompozy gives you a published week that still sounds like you.
It is an Apache-2.0 fine-tune of Alibaba's Qwen3-4B-Instruct-2507, released in October 2026, trained to write without common AI tics such as em dashes, "not X, but Y" phrasing, and lists of three. It runs locally — the 2.5 GB GGUF works in Ollama — and generates text only.
Yes, it is free under the base model's Apache-2.0 license. You can run the full bf16 weights with transformers, pull the 2.5 GB Q4_K_M GGUF into Ollama or llama.cpp, or apply the 126 MB LoRA adapter on top of Qwen3-4B-Instruct. It needs no system prompt.
On the author's own slop metric it dramatically cuts the measured tics and keeps factual answers correct, but the metric is self-reported with no blind human test, and he notes the voice can go flat, invent names and places, and run long. It is a solid drafting aid you should still edit and fact-check, not a finished-copy generator.
Yes, upstream. Draft a de-slopped paragraph locally, then paste it into Kompozy as a seed; Kompozy turns it into a captioned short, a carousel, quote graphics, per-platform posts, a blog, and a newsletter — all held to one voice by a Persona Brief — and publishes them across the eight social platforms plus blog and email. Kompozy's own copy runs on Claude and OpenAI, so the model is a drafting habit, not a plug-in.
No. qwen3-4b-deslop is a text-only model — it produces no images, video, carousels, or audio, and it does not schedule or publish anything. For finished, multi-format, published content you pair its drafts with a generation-and-publishing engine like Kompozy.