Erik Kannike has released qwen3-4b-deslop, a fine-tuned version of the Qwen3-4B-Instruct AI model that eliminates common AI writing tics such as em dashes and repetitive phrasing. The 2.5 GB model was trained on a MacBook Air and is available for local use via Ollama, marking a novel approach to improving AI writing style without sacrificing accuracy, according to 100toolkit.com.
Kannike trained the model on 161 answers and tested it on 40 prompts it had not seen before. The fine-tuning reduced the 'slop score'—a measure of unwanted writing habits—from 121 to 12.5 per 1,000 words. The model no longer uses em dashes and avoids the 'not X, but Y' construction and lists of three. It still answered all 30 factual questions correctly and matched requested genres and lengths as often as the original version, per 100toolkit.com.
This development addresses a common criticism of AI-generated text, which often contains stylistic quirks that make it identifiable as machine-produced. By removing these tics, qwen3-4b-deslop produces prose that 68% of the time scores within the range of human writing. The model’s voice is described as plain and flat, and while it still invents names and places, it represents progress in making AI writing more natural and less formulaic.
The model is downloadable as a 2.5 GB file and can be run locally using Ollama with the command 'ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M'. Kannike plans to apply the same fine-tuning process to larger models next, expanding the approach beyond the current 4B parameter size.