{"repo":"Firmamento-Technologies/TurboQuant","free":true,"listed":false,"github":"https://github.com/Firmamento-Technologies/TurboQuant","clone":"git clone https://github.com/Firmamento-Technologies/TurboQuant.git","description":"Near-optimal vector quantization from Google's ICLR 2026 paper — 95% recall, 5x compression, zero preprocessing, pure Python FAISS replacement","language":"Python","stars":20,"topics":["ann-search","approximate-nearest-neighbor","compression","deep-learning","embedding-compression","faiss","iclr-2026","information-retrieval","kv-cache","llm"],"license":"Apache-2.0","category":"machine-learning","readme_excerpt":"TurboQuant The only TurboQuant implementation for vector search — FAISS-compatible vector quantization library 180+ repos implemented Google's TurboQuant for KV cache compression. This is the only one built for vector similarity search. A pure Python implementation of the TurboQuant algorithm (Zandieh et al., ICLR 2026) for FAISS-compatible vector quantization . Compress embedding vectors by 5-8x with 95%+ recall , zero preprocessing , and no GPU required . Includes both brute-force ( TurboQuantIndex ) and sub-linear ANN search ( IVFTurboQuantIndex ). Why TurboQuant? Feature FAISS PQ ScaNN TurboQuant --------- ---------- ------- ---------------- Preprocessing K-means (minutes) Tree building (minutes) None (instant) ¹ Recall@10 60% 85% 95.3% Compression 8x 4x 5-8x Dependencies C++/CUDA C++/TensorFlow Pure Python/NumPy Theory guarantee None None 2.7x Shannon limit Training data needed Yes Yes No (data-oblivious) ¹ Query complexity O(N) brute-force O(log N) O(sqrt(N)) with IVF ¹ TurboQuantIndex : no preprocessing, no training data. IVFTurboQuantIndex : requires K-means training on representative data. Quick Start Installation Or from source: Basic Usage With IVF for Large Datasets With Sentence Transformers Low-Level Quantizer API Demo 8-second walkthrough: input vectors → random rotation → Lloyd-Max quantization → reconstruction quality. Full MP4 Performance How It Works TurboQuant implements a mathematically elegant compression scheme from Google Research (arXiv:2504.19874): A","default_branch":null,"files":null,"tree":[],"storefront":"/r/Firmamento-Technologies","claimed":false,"request_supported":{"post":"https://gitbuyer.com/r/Firmamento-Technologies/TurboQuant/request-supported","requests":0},"note":"indexed from public GitHub; nothing is for sale on this page. Clone it from GitHub. Paid listings live at /search."}