1. Qwen3-4B with the AI writing tics trained out
Best for: Local-LLM tinkerers and researchers who want a small model that writes short English prose without common assistant tics.
Pros
- Open Apache-2.0 weights: merged model, 126 MB LoRA adapter and four-neuron edit, plus Q4_K_M (2.5 GB) and Q8_0 GGUFs; runs locally with Ollama or llama.cpp
- Author's eval on 40 held-out prompts: slop score 121 → 12.5 per 1,000 words, em dashes 20.5 → 0, 30/30 factual quiz, genre/length fit 80% vs 78% for the base model, with no system prompt
- Transparent write-up of every training round, including failures (invented studies, ', which' chains) and how they were filtered out
Biggest catch
The author says the voice is flat, it still invents names and places, and it runs long (~200 words when asked for 150)
Skip it if: You want a hosted writing app, long-form, code or non-English output, or text that passes AI detectors.
Starting price: Free: open weights under Apache-2.0 on Hugging Face (bf16 merged model, LoRA adapter, Q4_K_M 2.5 GB and Q8_0 GGUF). Runs locally via Ollama, llama.cpp or transformers; no hosted version. (checked Oct 20, 2026)
Full Qwen3-4B with the AI writing tics trained out review · Qwen3-4B with the AI writing tics trained out alternatives