Glance: Unified 9B Backbone for RAG Generation and Reranking
October 28, 2026
Glance integrates reranking into a single 9B generator using a lightweight decision adapter and pointer head. This avoids redundant model residency and prefilling, achieving 0.741 nDCG@10 on SciFact with only 0.1 GiB of additional memory per decision skill.
HOW THIS AFFECTS YOU
●
builderYou can significantly reduce RAG infrastructure costs by using one model for both retrieval scoring and generation.
●
founderThis architecture enables more efficient, lower-latency RAG products on smaller GPU footprints.