A researcher has released Qwen3-4B-deslop, a small open-source model fine-tuned to write without common artificial intelligence writing patterns. According to the source, the model was trained on a MacBook Air using distillation techniques and addresses what the creator calls “AI slop”—recurring stylistic habits that mark text as machine-generated.
The model targets specific writing tics identified in a custom style guide: em dashes, “not X, but Y” constructions, lists of three items, rhetorical questions, signposts like “here’s the thing,” short fragments for emphasis, and words like “delve” and “seamless.” The creator developed a scoring system that weights each habit and counts occurrences per 1,000 words. According to the source, the original Qwen3-4B model averaged a score of 121, while human-written prose typically scores close to zero.
The training process involved three rounds of refinement. The first round used distillation, where a teacher model trained on 288 prompts generated answers that a smaller LoRA adapter learned to reproduce, achieving a score of about 20. The second round incorporated neuron steering and targeted patches, though this produced problematic outputs including text that wandered off-topic and invented citations like a fake study from the University of Technology Sydney. The final round added task-specific constraints: opinion pieces must state a position, personal reflections must use first person, and the model must avoid inventing names, places, or statistics.
The released model achieved a score of 12.5 with 80% of answers falling within human-writing ranges, according to the source. It no longer uses connective phrases like “, which” at abnormal rates and rarely invents false studies or percentages.
However, the model has limitations. It produces flat prose consisting of short declarative sentences, and personal writing reads like diary entries. The creator notes it still invents specific details like fictional coffee shops and names. Answers often exceed requested lengths, and testing has been limited to short English prose in specific genres; code, longer conversations, other languages, and tool use have not been thoroughly evaluated.
The 2.5 GB quantized model runs locally via Ollama or llama.cpp, requiring no system prompt. The full-precision weights are available through the transformers library. The source also notes planned expansions to larger models and blind testing with human readers.
Key facts
- The model targets specific AI writing habits including em dashes, “not X, but Y” phrases, lists of three, and filler words like “delve”
- Original Qwen3-4B scored 121 on the AI-writing-tics scale; the fine-tuned version scores 12.5
- Training used three rounds of distillation, with the final round adding task-specific constraints to ensure opinion pieces take positions and personal writing uses first person
- The 2.5 GB model runs locally on a MacBook Air via Ollama or llama.cpp without requiring a system prompt
- 80% of the model’s answers fall within human-writing ranges, though it still occasionally invents fictional names, places, and details