- Qwen3-4B model was further trained to reduce writing habits that look AI-generated. According to the creator, it uses fewer dashes, fewer 'not X but Y' sentences, and fewer lists grouped in threes, and it can be used without a separate system prompt.
- The creator's own rule-based detector scored it at 1000 per word, dropping from 121 for the original model to 12.5. This score is based on the creator's regex criteria, and no blind test with readers has been done yet.
- The creator said the training data included 161 answers generated on a MacBook Air. The creator also reported that it got all 30 answers right in a quiz of 30 factual questions, and this is the creator's own measurement.
- The original post says it can be run in Ollama by downloading the Q4_K_M quantized file (about 2.5GB). The original weights and the LoRA adapter are also published.
O: Marketers who often draft blog posts or LinkedIn posts could ask this model and the original model for the same topic through Ollama, then compare for themselves how much the dashes and lists grouped in threes have decreased.
X: The score 12.5 comes from a regex detector built by the creator, and there is no blind test with readers, so you need to check yourself whether the text actually reads less mechanically.