TurboQuant
Rotation-based vector quantization algorithm from Google Research. Zandieh et al., 2026 (arXiv:2504.19874). Now ships in Qdrant 1.18 with production extensions. Part of the Vector Quantization family; in the same design space as RaBitQ.
Core Algorithm
- Random orthogonal rotation — every vector is rotated; redistributes per-coordinate variance evenly so coordinates approximate N(0, 1) post-rotation
- Coordinate-wise quantization — fixed Lloyd-Max codebook for the standard normal distribution; universal — no per-dataset training, no calibration set, no stored codebooks; one lookup table works across all embeddings and dimensionalities
- Scoring — reconstruct dot product directly from codebook indices; orthogonal rotation preserves dot products and L2 distances
Two Paper Variants
| Variant | Description | Trade-off |
|---|---|---|
| MSE | Literal recipe; codebook lookup only | Symmetric: any pair of stored vectors can be scored from indices alone |
| PROD | Adds QJL random projection to cancel per-vector length bias | Better bias correction; splits bit budget between codebook and projection |
Qdrant ships MSE: symmetric scoring is required for HNSW graph construction; the length bias is fixed more cheaply via RaBitQ renormalization.
Compression Operating Points
| Variant | Bits/dim | Compression |
|---|---|---|
| 4-bit | 4 | 8× vs float32 |
| 2-bit | 2 | 16× |
| 1.5-bit | 1.5 | ~21× |
| 1-bit | 1 | 32× |
Compression vs. Storage Format
The 4-bit operating point above can be deployed two different ways, and the difference is larger than the shared bit-width suggests.
As a quantization mode, the compressed vectors accelerate search while float32 originals remain on disk; top candidates are rescored against the originals, which recovers most of the recall lost to compression. Total cost is 36 bits per coordinate — original plus code.
As a storage format — Qdrant’s Datatype.TURBO4, added in 1.19 — only the 4-bit code is written. Storage drops to 4 bits per coordinate, and the rescoring stage disappears with the originals, so quantization error becomes permanent rather than a first-pass approximation. This is the operating point where quantization stops being free.
See Qdrant 1.19 - Turbo4 Datatype and Memory Tiers.
Performance vs. Alternatives (Qdrant benchmarks, 10 datasets)
- TQ 4-bit vs SQ (4×): competitive — within 2 pp on 9/10 datasets; beats SQ on 3/10 (up to +4.6 pp); half the storage
- TQ 2-bit vs BQ 2-bit: +9–24 pp recall on every dataset
- TQ 1-bit vs BQ 1-bit: +9–21 pp recall on every dataset; still beats asymmetric BQ
Qdrant’s Production Extensions
Beyond the paper: Qdrant adds length renormalization from RaBitQ (per-vector scalar, 4 bytes), per-coordinate anisotropy compensation (P-Square quantile calibration per segment, free at query time), full L2/dot/cosine support (stores original L2 norm), and SIMD kernels (AVX-VNNI/AVX-512/NEON).
vs. Related Algorithms
| Algorithm | Quantization approach | Key difference |
|---|---|---|
| RaBitQ | Rotation + binarization | Length rescaling + bit-plane scoring; complements TurboQuant |
| BBQ / OSQ | Binary + centering | Elasticsearch; uniform integer grid; SIMD-optimized |
| Product Quantization | Sub-vector codebooks | Per-dataset trained; high compression, higher build cost |
Articles
- TurboQuant in Qdrant — Qdrant 1.18 implementation, production extensions, full benchmarks
- Qdrant 1.19 - Turbo4 Datatype and Memory Tiers — turbo4 as a storage datatype, without retained originals
- Elasticsearch BBQ Optimized Scalar Quantization vs TurboQuant — Elastic’s OSQ vs TurboQuant benchmarks
People
- Mohamed Arbi Nsibi — turbo4 storage datatype writeup
- Ivan Pleshkov — Qdrant; TurboQuant implementation
- Thomas Veasey — Elastic; OSQ/BBQ comparison benchmarks
Related Concepts
- Vector Quantization — parent concept
- Scalar Quantization — baseline at 4× compression; TurboQuant 4-bit achieves 8× at comparable recall
- Binary Quantization — baseline at 16–32×; TurboQuant beats BQ by 9–24 pp at same storage
- RaBitQ — companion rotation-based algorithm; contributes length renormalization and 1-bit scoring
- BBQ — Elasticsearch’s approach
- HNSW — TurboQuant’s MSE variant chosen for HNSW compatibility (symmetric scoring)
- Dense Vector Retrieval
- ANN
Mathematical Deep Dive
- The Mathematics of Google’s TurboQuant — Prateek Chandra Jha; first-principles derivation of PolarQuant + QJL; Beta distribution, Lloyd-Max quantizer, Johnson-Lindenstrauss lemma; within 2.7x of information-theoretic limit