Pageshift's PageStorm writes full books in one pass, setting new benchmark for creative AI

Pageshift Entertainment has introduced PageStorm Research Preview, a language model purpose-built for generating complete books in a single pass. Unlike conventional LLMs that require iterative prompting or chapter-by-chapter assembly, PageStorm handles book-length sequences — from tens to hundreds of thousands of tokens — in one forward pass. The model is trained on the LongPage Dataset, which the team released six months ago as a benchmark for creative writing at book scale. According to the accompanying preprint (arXiv:2605.17064), PageStorm achieves coherent plot development, consistent character arcs, and narrative structure across entire novels. The model's architecture and training objectives focus on preserving long-range dependencies, a known challenge for most autoregressive models. Pageshift highlights that PageStorm can produce both fiction and non-fiction book drafts, giving authors a starting point that requires only light editing rather than heavy rewriting.

PageStorm is available now as a research preview on HuggingFace, with weights and inference code open for experimentation. The team notes that while the model is not yet production-ready for commercial publishing, it represents a leap forward in generative AI for creative writing. The release includes configurable parameters for genre, length, and style. For AI researchers, PageStorm offers a testbed for studying coherence in ultra-long texts. For writers, it reduces the friction of starting a novel from scratch. Pageshift plans to iterate based on community feedback, aiming toward a full release later this year.