1-Billion-Token Context Window. It's Absolutely Unreal.

Abstract

An intentionally provocative engineering note accompanying the open-source biggie-kun project. It describes a request-scoped lexical retrieval system presented through a chat-completions interface as a nominal one-billion-token context window. The paper discusses the implementation, retrieval loop, scaling strategy, and limitations of using agentic retrieval instead of native long-context attention. This work is a preprint and has not been peer reviewed. Source code: https://github.com/tnfssc/biggie-kun Live project: https://biggie-kun-vps.sharath.page