GitHub radar
Qwen3-4B-deslop: a local model that drafts text without AI tics
This is a small AI model retrained to drop the habits that make text read as AI-written: em dashes, three-item lists, "not X, but Y." Run it on your own laptop and get a draft that reads less like AI.
A retrained version of Qwen3-4B: typical AI clichés are trained out of the model's weights directly, with no system prompt (the instruction given before every reply) and no extra code.
When it helps
Useful for a quick draft without obvious AI-style tells, or to check your own text for them, right on your laptop. It's weak at complex tasks and agent work: the voice is flat and it can invent facts.
Pros
- Tiny — 2.5 GB, runs on an ordinary laptop with no GPU
- Drops em dashes, three-item lists, and "not X, but Y" phrasing
- Answers factual and math test questions correctly
- Installs with one command in Ollama
Cons
- Voice comes out flat, almost no personality
- Sometimes invents details, numbers, and names
- Can loop on one phrase in long answers
How to set it up — step by step
- 1Install Ollama (a tool for running AI models on your computer) if you don't have it.
- 2Open a terminal and run: ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M
- 3Give the model one of your drafts and ask it to rewrite it without AI tics.
- 4Send your agent the text from the block below if you work through an AI agent (Claude Code, Codex).
- 5Compare its answer with a regular model's answer and judge which sounds more natural.
- 6Check facts and numbers in the answer by hand — the model can make them up.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the Qwen3-4B-deslop model locally from Hugging Face: https://huggingface.co/erikkannike/qwen3-4b-deslop-v4 — use the Ollama command from the README (ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M), run one of my own working texts through it, and show me its answer next to a regular model's answer so I can compare the style.
Runs on a regular laptop, no GPU needed: the quantized version is 2.5 GB and starts with one Ollama command.
Open on Hugging Face▌ More finds