Google Research Unveils TurboQuant Memory Compression

Google Research developed TurboQuant, a memory-compression algorithm for AI inference, and will present results at ICLR 2026 next month. The method uses vector quantization—including PolarQuant and a QJL training/optimization approach—to shrink KV cache runtime memory by at least six times without degrading performance. If validated, TurboQuant could significantly lower inference memory footprints and operational costs for large models.