LongWriter

Why we included this project

LongWriter is for anyone who has watched a model start repeating itself or trail off entirely after a few thousand tokens. It shows how to get LLMs to produce coherent 10,000+ word outputs, and it ships enough to try that out right away: two ready-to-load open weights (one based on GLM-4-9B, one on Llama-3.1-8B) and a single inference script that works with either transformers or vLLM. The repo also opens up the whole training recipe, which matters if you want to reproduce the results rather than just use them. That means the AgentWrite pipeline, which breaks a long request into an outline of sub-tasks and uses it to build training data, plus the fine-tuning code and two evaluation benchmarks that measure output length separately from quality. If supervised training is not your route, the newer LongWriter-Zero variant uses pure RL with no synthetic or annotated data and reportedly beats even much larger models on long-form writing. Teams producing long drafts, reports, or fiction from fine-tuned or on-premise models will get a clearer picture of what it takes to make a model actually finish a long piece of writing.