Skip to main content
Back

Qdrant 1.19 Shrinks Vector Database Storage by 9x With 4-Bit Compression

Qdrant 1.19 ships a 4-bit TurboQuant storage datatype that cuts vector disk usage 9x, plus unified memory tiers and slice filters.

Repo, Full Access
  • Qdrant 1.19 ships a Turbo4 storage datatype that stores vectors as 4-bit only, cutting disk usage roughly 9x.
  • A single memory parameter with pinned, cached, and cold tiers replaces the old per-component storage flags.
  • Per-tenant IDF statistics fix BM25 scoring in multi-tenant collections where tenants use different vocabularies.
  • New filter primitives add prefix matching on keyword indexes and slice conditions for deterministic partitioning.
  • Legacy /search, /recommend, and /discover endpoints are removed; migrate to the unified /query API before upgrading.
  • Web UI adds live resharding progress, a WebGL2 collection visualizer, and interactive payload index management via the v1.19.0 release.

The team behind the Rust-based vector database Qdrant just cut the on-disk footprint of vector collections by roughly an order of magnitude. Their 1.19 release introduces a new primary storage datatype built on 4-bit TurboQuant compression, along with a reworked memory configuration model, per-tenant BM25 statistics, and filter conditions that make parallel scans and reproducible sampling straightforward.

If you are running large collections and watching your storage bill, the headline change is worth a close look: Qdrant now offers a storage-native 4-bit format that keeps only the compressed vectors on disk, alongside unified memory tier configuration, replica read affinity, and full-text search improvements.

From two copies to one

TurboQuant landed in Qdrant 1.18 as a quantization technique, meaning a compressed shadow copy of your vectors used to speed up search. Qdrant kept the original full-precision floats on disk and used the smaller version during HNSW graph traversal, then rescored the top candidates against the originals to preserve recall. Accuracy was good, but storage doubled up.

Version 1.19 flips this into a storage-native format. The new Turbo4 datatype uses 4-bit TurboQuant compression as the only representation of your vectors, with no full-precision copy retained. The blog post pegs the savings at roughly nine times less storage per coordinate compared to keeping both representations, with proportionally larger wins for ColBERT-style multi-vector collections used in late interaction retrieval.