Qwen3-4B Deslop: Stripping Artificial Prose Quirks via Targeted Neuron Edits and LoRA
An independent researcher released a modified Qwen3-4B model that removes robotic AI writing clichés by flipping four specific MLP layer weights and fine-tuning with a rank-16 LoRA adapter. It runs locally in 2.5 GB of VRAM.
Why it matters
Run this 4B model locally in Ollama to generate natural, punchy prose without typical LLM tropes like overused metaphors, pompous tone, and repetitive transitions.
TL;DR
- 01Sign-flipping four MLP neurons across layers 28, 32, and 33 suppresses typical AI writing mannerisms.
- 02The model produces punchy, short declarative sentences but struggles with exact word count limits.
- 03Q4_K_M GGUF weights run comfortably on consumer laptops within 2.5 GB of VRAM.
Key facts
- Base model
- Qwen3-4B-Instruct-2507 in bf16
- Quantized size
- 2.5 GB (Q4_K_M GGUF)
- LoRA config
- Rank-16 on all attention and MLP projections, 2 epochs
- Neuron modification
- 4 MLP neurons negated across layers 28, 32, and 33
The Deslop Architecture
The project modifies Qwen3-4B-Instruct-2507 (bf16) using a dual-intervention strategy trained on an M2 MacBook Air:
1. Neuron Edit (175 bytes): Flips the weight sign of four MLP neurons by negating columns in their layers' down-projections: layer 28 (neuron 1121), layer 32 (neurons 4852 and 5418), and layer 33 (neuron 348). 2. LoRA Adapter (126 MB): A rank-16 adapter applied across all attention and MLP projections, trained for two epochs on bare user prompts.
Running Locally in Ollama and Transformers
The quantized checkpoint runs locally via standard Ollama or llama.cpp setups, consuming approximately 2.5 GB of RAM.
Try it in 2 minutes
# Run the deslopped Qwen3-4B model locally in Ollama
ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M
# Or run via llama-cli directly
llama-cli -hf erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_Mbash
✓ When to use
- Use for drafting marketing posts, personal blog essays, and technical explainers without standard LLM buzzwords.
- Use on resource-constrained developer laptops requiring offline local inference under 3 GB RAM.
✕ When NOT to use
- Do not use as a code completion engine or autonomous programming agent.
- Do not rely on it for strict word-count compliance, as it consistently overshoots length targets by ~30%.
What to do today
- Pull and test the Q4_K_M GGUF build via Ollama for local documentation and draft generation.
- Evaluate your output against strict style regexes to measure AI drift in technical prose.
- Avoid using this checkpoint for agentic tool use or coding tasks without running domain benchmarks.
What the community says
“The most difficult thing in doing it yourself is knowing what to ask for. I can ask my model to talk like a caveman or like a Baltic pirate. But asking for something more nuanced... requires the correct lingo”
“The tell tell sign for me is that it's still incredibly hard to read. AI tend to write in a style that I'm unable to easily read... text simply doesn't register.”
“It's definitely a slight improvement, but it is still rather obvious that the output is AI-written :(”