Skip to content

TurboQuant

Data-oblivious online quantization with near-optimal distortion (ICLR 2026).

TurboQuant compresses vectors to 1, 2, 4, or 8 bits per dimension without a training phase. At 4-bit on 768-dimensional embeddings, it achieves a 7.5× memory reduction while preserving > 90% recall.

When to use

  • Extreme memory pressure: 10× or more vectors than you can fit in RAM at full precision
  • Billion-scale collections where IVF_TURBOQUANT enables sub-linear search
  • Cost-sensitive deployments: fewer replicas, smaller nodes

Index types

TurboQuant

Pure quantization over a flat scan. Compresses vectors but still evaluates every one.

client.create_collection(
    "compressed",
    dimension=768,
    index_type="turboquant",
    # bits=4,  # 1, 2, 4, or 8
)

IVF_TURBOQUANT

Combines IVF partitioning with TurboQuant. The coarse quantizer skips irrelevant clusters; TurboQuant compresses what's scanned. This is the default choice for AUTO at ≥ 1M vectors.

client.create_collection(
    "billion_scale",
    dimension=768,
    index_type="ivf_turboquant",
)

Compression vs recall

Rough guidance at dimension 768:

Bits Memory reduction Typical recall@10
8 4× > 97%
4 7.5× > 90%
2 15× 85–90%
1 30× 75–85%

Tune per workload — recall depends on data distribution.

How it works (briefly)

TurboQuant uses a data-oblivious randomized projection followed by per-dimension scalar quantization. Unlike product quantization, there is no codebook to train; quantization parameters are derived from the input vector at insert time. The distortion bound is near-optimal (within a constant factor of the information-theoretic lower bound for the given bit budget).

The reference: TurboQuant: Online Near-Optimal Quantization for Vectors (ICLR 2026).

Further reading