TurboQuant: Redefining AI efficiency with extreme compression

TurboQuant was introduced to be presented at ICLR 2026.

PolarQuant was presented as part of TurboQuant to be presented at AISTATS 2026.

Quantized Johnson-Lindenstrauss reduces vectors to one-bit signs using the Johnson-Lindenstrauss Transform reduces each vector number to a sign bit.

Experiments evaluated TurboQuant, QJL, and PolarQuant on LongBench, Needle In A Haystack, ZeroSCROLLS, RULER, and L-Eval tested on long-context benchmarks listed in the article.

Benchmarks used open-source models Gemma and Mistral in evaluations used open-source LLMs named in the article.