Choosing Affordable Language Models for Serial Fiction: Instruction Following, Creative Writing, and Story Memory

This preprint compares six affordable language models for AI-assisted serial fiction through 687 chapter-generation attempts. It evaluates instruction detail and formatting, chapter completion, selected event coverage, creative writing, character and dialogue portrayal, story-bible updates, history compression, retrieval, and guided continuation. The paper provides model-specific settings, a practical chapter contract, and cost examples for authors. The study separates instruction following from resemblance to a reference chapter and from literary quality. Findings come from one selected novel; literary assessments are exploratory AI readings rather than an independent human evaluation, and reliable long-novel or cross-genre performance is not established. Models studied: DeepSeek V4 Flash 0731, GLM 5.3 Flash, Xiaomi MiMo-V2.5, Tencent Hy3, Google Gemini 2.5 Flash Lite, and DeepSeek V3.2. Open-source writing harnessThe companion writing harness is available under the MIT license at https://github.com/vamshibobby/webnovel-harness. Authors can run it locally with their own OpenRouter API key and choose compatible models to apply the paper's findings. The repository documents prefix/prompt-cache reuse, provider pinning, automatic chapter summarization, story-bible management through tool calls, retrieval of earlier chapters, character design and consistency tools, configurable generation settings, and usage and cost tracking. These features help authors manage continuity and generation costs; model compatibility, tool support, and caching savings depend on the selected model and serving provider. Setup instructions and the local web interface are included in the repository. The repository is a companion implementation, and its capabilities should not be read as additional experimental findings of this paper.