VQ-bench

a composable vector quantization framework

VQ-bench is an open-source benchmark and framework for vector quantization. It implements several algorithmic primitives and builds quantizers by composing these primitives. It evaluates the performance of existing quantizers on several size and quality metrics, and we publish the results on this page.

Quantizers

Each method evaluated in VQ-bench, with a link to its source.

QuantizerPaperYear

MinMax (per-vector) and Scalar (per-dimension) are uniform min–max quantization baselines with no dedicated reference.

Datasets
DatasetVectorsDimensionCalibration queries
ArXiv1,344,6437680
CCNews495,3287680
Yahoo677,3053840
COCO282,360768114,054
LAION1,000,448512999,448

All datasets are from VIBE (Jääsääri et al. 2025) and use inner-product similarity. Calibration queries is the number of available sample queries for training.

Metrics
MetricWhat it measuresBetter
KV-cache values / signal compression
reconstruction MSEaverage squared distance between the true vector and the reconstructed vectorlower
reconstruction biassquared norm of the average residual vector (true vector − approximated vector)lower
Model weights compression
score MSEaverage squared error of the inner-product estimates between queries and candidateslower
score biasbias of the inner-product estimates→ 0
Vector search / recommendations
recall@kfraction of the true top-k neighbors contained in the approximated top-khigher
SOS@kratio of summed top-k approximated scores and summed true top-k scoreshigher
exp-SOS@k (τ)SOS@k over exponentiated scores exp(score / τ), at temperature τhigher
KV-cache keys
attention KL (τ)KL-divergence between the softmax distributions induced by true and approximate candidate scores, at temperature τlower
attention TV (τ)TV-distance between the softmax distributions, at temperature τlower

We also report size and efficiency metrics (bits per dimension, peak / encode memory, encode / scoring time). All time and memory statistics were measured on an Apple M2 Pro with 16GB RAM using 6 threads.

VQ-bench is actively maintained by Amir Ingber, Edo Liberty (Pinecone), and Ashwin Padaki (UPenn).


Benchmarking Results

VQ-bench