Qwen3-4B-Deslop Cuts Its Slop Score From 121 to 12.5 on a MacBook Air, and Two of 40 Answers Came Out Damaged

A developer trained a 4-billion-parameter model on an M2 MacBook Air to stop writing like a chatbot, and its author's own regex-based score fell from 121 to 12.5 slop points per 1,000 words. The Hugging Face card for the model also says 2 of 40 test answers, 5%, were flagged as damaged, a figure the launch blog post does not give.

The release, Qwen3-4B-deslop, was posted to Hacker News on Monday as a Show HN by Erik Markus Kannike of 100toolkit.com. It had 136 points and 101 comments by Monday evening.

How it was trained

The base is Alibaba's Qwen3-4B-Instruct-2507, an Apache-2.0 model with 4.0B parameters and a 262,144-token native context window. Kannike's pipeline had a teacher model answer 288 prompts, then kept only answers that passed style and task-fit checks. The released version, the post says, was trained on 161 kept answers.

A rank-16 LoRA adapter learned the style from the bare prompt, over two epochs on the laptop. A separate edit flips the sign of four MLP neurons across layers 28, 32 and 33. The adapter is 126 MB and the neuron edit is 175 bytes, and the post says they must be used together.

The post says an earlier round overused ", which" joins and invented studies, and round three added task-fit checks.

The numbers, all from the author

Every result below comes from 40 held-out prompts with one sample each, scored by the author.

Measure	Original	Deslop v4
Slop score per 1,000 words	121	12.5
Em dashes per 1,000 words	20.5	0
Answers in the human range	0%	68%
Genre and length fit	78%	80%

1. Qwen3-4B-Instruct-2507121 points
2. Qwen3-4B-deslop v412.5 points

Source: Erik Markus Kannike, 100toolkit.com blog post, accessed 2026-10-19

The post says human writing usually scores under 15 on this scale. The score is a regex-based style guide, and the post says no blind reader study has been done. It also says scores shift between runs: one model measured 10.2 as an adapter and 12.5 as merged weights.

What the model card adds

The Hugging Face card lists the licence as Apache-2.0, 4B parameters in BF16 and no gating. The blog post gives no licence for the release or the adapter, so the card is the only place the term appears.

It also reports factual and arithmetic accuracy rising from 97% to 100%, where the post says 30 of 30 factual questions were answered correctly. Drift on human-written prose is 0.061 nats per token, against 0 for the original. And the card flags 2 of 40 answers as damaged.

The card showed 31 downloads in the last month and 6 likes when fetched.

Limits the author lists

The post says the voice is flat, personal pieces read like diary entries, and the model often invents names and places, such as a coffee shop with a named owner. It tends to write longer than asked.

Only short English prose in four genres was tested. Code, long conversations, other languages and tool use are untested, the post says.

The Ollama build is a 2.5 GB Q4_K_M file, run with ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M. A neuron edit applied by hand to a different runtime would need the 175-byte file as well as the adapter.

Related coverage of small open models: Liquid's LFM2.5-24B-A2B, whose licence ends at $10 million in revenue, and MiniCPM-V 4.7.

Kannike has not published the 40 held-out prompts or the detector's regex list in the post, so the 12.5 cannot be reproduced from it.