New Vector Index Cuts Storage Needs by 97 Percent
The LEANN index enables researchers to run large-scale vector searches with significantly reduced hardware memory requirements.
Updated on Oct. 27, 2026 in Artificial Intelligence
Live Poll
Is it worth sacrificing search speed to significantly reduce the storage footprint of your local AI?
Researchers at UC Berkeley have released LEANN, an open-source vector index that drastically reduces memory overhead for large datasets. This research-stage system optimizes storage by recomputing embeddings at query time instead of maintaining massive on-disk indices.
Why it matters
Vector indexes often balloon in size beyond the raw data they manage, creating significant storage bottlenecks for AI applications. LEANN demonstrates a method to trade off retrieval latency for a massive reduction in memory footprint, offering a path for running complex search tasks on resource-constrained hardware.
LEANN achieves a 97% reduction in storage, shrinking a 201 GB Wikipedia index down to 6 GB. While it requires 2.48 seconds for retrieval on NQ datasets compared to 0.05 seconds for standard HNSW (Hierarchical Navigable Small World) graphs, it maintains 90% recall at 25.5% downstream accuracy.
The players
UC Berkeley
A top-tier public research university known for pioneering AI infrastructure and large-scale data systems.
The details
LEANN achieves its compression by using a pruned graph—a map of data points with redundant connections removed—and a product quantization table, which is a lookup system that approximates vectors. Instead of saving every embedding to disk, the system recomputes the necessary nodes on the search path at query time. This method also uses dynamic batching to manage processing tasks and relies on high-degree preserving pruning to ensure the most relevant search paths remain intact for retrieval.
Timeline
2026: LEANN won the Best Research Paper award at MLSys 2026.
The Tech Race
The system marks a departure from the industry trend of increasing memory allocation to maximize retrieval speed. By winning Best Research Paper at MLSys 2026, it challenges the dominance of HNSW indexes that prioritize sub-millisecond response times at the cost of massive disk usage.
Developers running local AI pipelines on platforms including Ubuntu, Arch, WSL, macOS, and Windows can currently experiment with the code on GitHub. Compatibility with Ollama and Hugging Face means it can be integrated into existing LLM workflows that require compact, efficient retrieval systems.
The takeaway
LEANN provides a novel way to scale AI search capabilities on consumer hardware by prioritizing memory efficiency over raw speed. Watch for future benchmarks evaluating its performance against larger, real-world datasets to see if it remains viable as a primary index for high-traffic applications.
Further reading
Explore more developments in Artificial Intelligence to understand how research-stage vector compression influences model deployment.
More information
Access the source code and implementation details at the LEANN open-source vector index repository.
Source note: This article includes information reported by MarkTechPost.
Live Poll
Is it worth sacrificing search speed to significantly reduce the storage footprint of your local AI?
