qwen3-4b-deslop review (2026): an honest verdict on the Qwen3-4B fine-tune trained to kill AI writing tics — slop-score numbers, voice, and real limits.
qwen3-4b-deslop is a genuinely interesting hobby fine-tune: on the author's own slop metric it slashes the measurable AI tics — em dashes to zero, lists of three and "not X, but Y" way down — while keeping factual answers correct. For a privacy-minded tinkerer who drafts locally and wants plainer prose, it is a free, Apache-2.0, laptop-sized win. The honest limits are big: the slop score is self-reported with no blind human test, the voice goes flat and sometimes invents names and places, answers run long, and it is only a drafting brain — it makes no finished content and publishes nothing. Score it as a promising experiment, not a production copywriter.
qwen3-4b-deslop is a fine-tune of Alibaba's small open model Qwen3-4B-Instruct-2507, released by developer Erik Markus Kannike and posted as a Show HN on October 19, 2026. Its one job is to write prose that does not read as AI-generated — specifically to avoid em dashes, "not X, but Y" constructions, and lists of three, the habits the author treats as the clearest tells. This review scores it as it stands on 2026-10-19.
The appeal is real. Most "make it sound human" advice is prompt-wrangling you redo every session; this bakes the behavior into the weights, ships under Apache-2.0, and runs on a MacBook Air as a 2.5 GB quantized file. If you have ever deleted the same three em dashes out of every draft, the premise lands.
The skepticism is also real, and the author invites it. The headline numbers come from his own regex-based "slop score" with no blind human test behind them, the base is only a 4B model, and the card lists failure modes plainly: flat, diary-like voice, invented names and places, and answers that overrun the requested length. One intermediate version even learned to write opinion pieces with no opinion and to fabricate studies.
Disclosure: I build Kompozy, a content generation and publishing engine. The overlap with this project is narrow — we both care about output that does not sound robotic — so I have scored the model on its own terms as a local writing tool, not against a job it never set out to do. Short version: a clever, honest, free experiment that is strong at the narrow thing it measures and clearly unfinished at everything around it.
The model is Qwen3-4B-Instruct-2507 with a rank-16 LoRA adapter trained by distillation: a teacher model answered 288 prompts, answers that still showed the tics or missed the task were filtered out, and the adapter learned to reproduce the clean survivors from the bare prompt. It took three rounds on an M2 MacBook Air, with a task-fit filter added in round three to stop an earlier version from writing contentless opinion pieces and inventing studies. The release also includes an unusual 175-byte "neuron edit" that flips the sign of four MLP neurons across layers 28, 32, and 33, which the author credits with removing about 70% of the "not X, but Y" pattern on its own. It is distributed three ways on Hugging Face under the base model's Apache-2.0 license: full bf16 weights for transformers, a 2.5 GB Q4_K_M GGUF to run in Ollama or llama.cpp, and the 126 MB LoRA adapter (which expects the neuron edit active, since it was trained that way). It needs no system prompt or special scaffolding. What it is not: it generates only text, in a chat or API call, with no images, no video, no scheduling, and no publishing. It is a drafting model, full stop.
This fits a specific person: a developer or writer comfortable running a local model who wants plainer first drafts without sending every idea to a hosted API. If you draft in Ollama, value privacy, and are tired of hand-deleting AI tics, it is a free and pleasant upgrade over the stock 4B, and the open method makes it a good study for anyone curious about how de-slopping actually works. It is a poor fit for anyone who needs reliable, publication-ready copy — the 4B base, the flat voice, the invented details, and the length drift all mean you still edit and fact-check everything — and for creators who need finished content across platforms, because this produces raw text only and ships nothing.
| Dimension | Score | Why |
|---|---|---|
| AI-tic reduction (on its own metric) | 4.2 / 5 | Em dashes to zero, lists of three and "not X, but Y" sharply down — the thing it set out to do, it does. |
| Human-sounding voice (real feel) | 3.0 / 5 | Plainer, but trades tics for a flat, diary-like tone; 68% "human-range" is self-scored, not blind-tested. |
| Factual accuracy | 3.8 / 5 | 30/30 on the author's factual set, but a 4B base still has shallow knowledge and invents names and places. |
| Task fit (genre & length) | 3.3 / 5 | About 80% on target genre/length; answers tend to run longer than requested. |
| Setup & ease of use | 3.8 / 5 | 2.5 GB GGUF runs in Ollama with no system prompt; the LoRA path needs the paired neuron edit. |
| Versatility | 2.5 / 5 | Only short English prose in four genres tested — code, other languages, and tool use are untested. |
| Transparency & method | 4.6 / 5 | Open weights, documented pipeline, published metrics, and an unusually honest list of failure modes. |
| Value (free, open, local) | 4.5 / 5 | Apache-2.0, free, laptop-sized — excellent value for a drafting aid if the base model suits you. |
| Finished content & publishing | 1.0 / 5 | None — it is a text model with no media generation, scheduling, or distribution whatsoever. |
There is no price to analyze in the usual sense: the model is free and open under the Apache-2.0 license, with the weights, the GGUF, and the LoRA adapter all downloadable from Hugging Face. The only cost is the hardware to run it, and the author's own demonstration — training it on an M2 MacBook Air and quantizing to 2.5 GB — shows how low that bar is. For inference, any machine that can run a small quantized model in Ollama will do.
That makes the value proposition simple and strong for the right person: zero marginal cost per word, full local privacy, and no vendor lock-in. Compared with paying per token for a hosted frontier model and then prompting it to sound human every session, a free local model that bakes the behavior in is attractive if you are already set up to run one.
The cost that does not show up on a pricing page is your time. Because the output still needs editing and fact-checking, and because it makes only raw text, the real expense is the manual work of turning a de-slopped paragraph into finished, on-brand, published content. That is the gap a generation-and-publishing engine fills, and it is the right way to think about "cost" here — not dollars for the model, but hours for everything the model does not do.
| Use case | Fit | Why |
|---|---|---|
| Private, local drafting of plain prose | Strong | Free, laptop-sized, and purpose-built to drop the tics that mark a draft as AI-written. |
| Studying how de-slopping / anti-AI-tell works | Strong | Open weights, documented distillation pipeline, published metrics, and a novel neuron edit make it a great teaching case. |
| First-draft blog or post copy you will edit | OK | A reasonable starting point, but you must fact-check, add a point of view, and trim the length drift. |
| Reliable, publication-ready copy | Weak | A 4B base that invents names and places and goes flat is not trustworthy as finished output. |
| Brand-voice consistency across many formats | Weak | It cleans one block of text; it has no mechanism to hold one voice across captions, blogs, and carousels. |
| Producing finished video, images, or carousels | Weak | It generates text only — no media of any kind. |
| Scheduling and publishing content | Weak | There is no distribution layer; the output is a string you then have to place and post yourself. |
Be clear about the category difference, because it is the whole story. qwen3-4b-deslop is a drafting brain that outputs a cleaner block of text; Kompozy is a content operation that outputs finished, on-brand media across platforms. They are not competitors so much as different stages of the same job — the model can sit upstream of Kompozy as a place you draft, and Kompozy is where that draft becomes a captioned short, a carousel, a blog, and a newsletter that actually get published.
Where the two genuinely rhyme is the goal: output that does not sound robotic. This project trains the tics out of one 4B model for short English prose; Kompozy pursues the same end structurally, governing every text format with a Persona Brief plus banned-word filters so the robotic phrasing is screened at generation and one voice holds across captions, Text Posts, blogs, and newsletters — not just in the single box where you tested it. Kompozy's copy runs on Claude and OpenAI rather than this model, so the comparison is about approach, not integration. If you want a free local toy that drafts plainer paragraphs, this is a fun one; if you want an un-robotic voice shipped as a week of multi-format content across the eight social platforms plus blog and email, that is a different tool, and it is the one Kompozy is.
It is an Apache-2.0 fine-tune of Alibaba's Qwen3-4B-Instruct-2507, released by Erik Markus Kannike in October 2026, trained to write without common AI tics — em dashes, "not X, but Y," and lists of three. It is distributed on Hugging Face as full weights, a 2.5 GB GGUF for Ollama, and a LoRA adapter.
For a privacy-minded tinkerer who drafts locally, yes — it is free, runs on a laptop, and measurably reduces AI tics while keeping factual answers correct. For anyone needing reliable, publication-ready copy it is not enough on its own: the voice goes flat, it invents names and places, answers run long, and it generates text only, so you still edit, fact-check, and build the finished content yourself.
On the author's own regex "slop score" the improvement is dramatic — 121 to 12.5 per 1,000 words, with em dashes eliminated — and he reports 68% of answers in a "human range." But that metric is self-reported with no blind human study, and the model trades the tics for a flat tone. Call it less obviously AI, not reliably human.
Nothing. It is free and open under the Apache-2.0 license, with weights, a GGUF, and a LoRA adapter all downloadable. The only cost is the hardware to run it — and a 2.5 GB quantized model runs on modest machines — plus the time to edit its output and turn it into finished content.
The slop metric is unvalidated by blind testing; the voice is flat and sometimes invents names and places; answers overrun the requested length; the 4B base has shallow knowledge; and only short English prose in four genres was tested, so code, other languages, and tool use are unknown. It also produces raw text only — no media, scheduling, or publishing.
qwen3-4b-deslop is a local text model that outputs a cleaner paragraph. Kompozy is a generation-and-publishing engine that turns one idea into video, carousels, quote graphics, blogs, and newsletters — all held to one voice by a Persona Brief with banned-word filters — and then schedules and publishes them across the eight social platforms plus blog and email. The model drafts; Kompozy produces and distributes.
Not as a plug-in — Kompozy's copy runs on Claude and OpenAI. But you can use the model upstream: draft a de-slopped paragraph locally, then paste it into Kompozy's Quick Ingest as a seed, and Kompozy fans it into finished, on-brand formats and publishes them. The model is where you draft; Kompozy is where it ships.