In early May 2026 Qdrant shipped TurboQuant, a quantization layer that promises memory savings without destabilizing vector‑search recall. Traditional scalar quantization trims Float32 embeddings to uint8, cutting size by four but adding modest error; binary schemes push compression to 32× at the cost of noisy results. TurboQuant claims to keep geometry intact while still shrinking vectors for large‑scale deployments in practice.

TurboQuant’s trick is a random orthogonal rotation applied before any bits are allocated. The rotation spreads variance evenly across all dimensions, so a single pre‑computed codebook can compress the vector without per‑dimension tuning. This approach, derived from a Google Research paper presented at ICLR 2026, avoids the signal‑to‑noise bias of scalar and binary methods and requires no extra overhead.

The author benchmarked TurboQuant against scalar and binary quantization on datasets ranging from thousands to millions of vectors. Results showed comparable recall to scalar quantization while using roughly half the storage, and far more stable similarity scores than binary compression. For production workloads that balance latency, cost, and accuracy, TurboQuant now merits consideration as a default quantizer especially when handling high‑dimensional data.