LocalLLaMA hunts chapter-book prose models

The real story here is that serious local fiction writing still looks more like a workflow problem than a single-model breakthrough.

• –The original poster rejects APIs outright and describes using SillyTavern in context mode specifically to avoid the short-turn behavior common in instruct and RP-tuned models.
• –Early replies point to EQ-Bench’s creative-writing leaderboard and the UGI leaderboard, showing that prose-specific evals are becoming more important than generic chatbot benchmarks for this niche.
• –Recommendations in the thread lean toward heavier models such as GLM 4.6/4.7 and Qwen 3.5 27B or larger, with commenters arguing they avoid premature wrap-ups better than many mid-sized RP finetunes.
• –Experimental finetunes from DavidAU get praise for prose quality but also criticism for instability, which captures the current tradeoff in local writing models: style versus reliability.
• –The broader LocalLLaMA consensus from related discussions is that notebook-style prompting, raw text completion, and chapter-by-chapter steering matter almost as much as the base model itself.