Residual Quantization with Implicit Neural Codebooks

A commutative vector quantization method for KV cache compression that reduces FP16 cache size by 87.5% with 2-bit quantization and enables 1-bit quantization, allowing LLaMA-3.1 8B to run with 128K context on a single RTX 4090 GPU.

Advanced vector quantization technique that explores the Leech lattice's optimal sphere packing properties at 24 dimensions. Delivers state-of-the-art LLM quantization performance, outperforming recent methods like Quip#, QTIP, and PVQ for extreme vector compression.

Compression technique that converts full-precision vectors to binary representations, achieving 32x storage reduction while maintaining 90-95% recall for efficient large-scale vector search.

Compression method developed by Timescale researchers that improves on standard Binary Quantization, reducing vector memory footprint by 32x while maintaining high accuracy for filtered searches.

Elasticsearch and Lucene's implementation of RaBitQ algorithm for 1-bit vector quantization, renamed as BBQ. Provides 32x compression with asymptotically optimal error bounds, enabling efficient vector search at massive scale with minimal accuracy loss.

Advanced quantization technique that applies per-vector normalization and scalar quantization, adapting the quantization bounds individually for each vector. Achieves four-fold reduction in vector size while maintaining search accuracy with 26-37% overall memory footprint reduction.