UC Berkeley open-sources LEANN, cutting a 60 million-paragraph index from 201 GB to 6 GB

UC Berkeley’s Sky Computing Lab has released LEANN, an open-source vector index designed to reduce storage needs by avoiding the long-term storage of full vector embeddings and computing them at query time instead. According to the figures cited in the report, LEANN shrank the index size for a Wikipedia corpus containing 60 million paragraphs from 201 GB to 6 GB. The research paper behind the system won the Best Paper Award at MLSys 2026. LEANN is built on FAISS and keeps a compact graph structure while discarding full vectors. In the NQ benchmark, using the Qwen3-4B model on an RTX 4090, LEANN and HNSW both reached 25.5% downstream accuracy. The storage footprint, however, was far smaller for LEANN at 4 GB, compared with 188 GB for HNSW. That reduction came with slower retrieval: 2.48 seconds for LEANN versus 0.05 seconds for HNSW on the NQ task. The project is aimed at making semantic search and retrieval-augmented generation, or RAG, practical on personal laptops. A GitHub repository cited in the report also showed index reductions from 1.8 GB to 64 MB for 400,000 chat messages and from 130 MB to 6.4 MB for 38,000 browser records.