Skip to main content
Back

turbovec Shrinks 31 GB Vector Indexes to 4 GB and Beats FAISS

turbovec packs 10M vectors into 4GB of RAM using Google's training-free TurboQuant algorithm, beating FAISS on ARM with a one-line pip install

  • 16x compression: turbovec shrinks a 31GB float32 corpus of 10M vectors down to ~4GB using Google Research's TurboQuant algorithm.
  • Zero training required: Unlike FAISS PQ, no codebook training or data passes needed -- vectors are indexed instantly on arrival.
  • Faster than FAISS on ARM: Hand-written NEON kernels beat FAISS IndexPQFastScan by 12-20% on Apple Silicon; competitive on x86.
  • ICLR 2026 paper: Built on TurboQuant, proven to operate within 2.7x of the Shannon distortion limit at any bit-width.
  • Drop-in Python install: pip install turbovec with native integrations for LangChain, LlamaIndex, and Haystack.
  • Limitations: Flat-scan only (no HNSW), no GPU support, and 2-bit multi-threaded configs can trail FAISS by 2-4% on x86.

turbovec is a new open-source vector index written in Rust with Python bindings that compresses embedding corpora by up to 16x and searches them faster than FAISS on ARM hardware. The headline number is hard to ignore: a 10 million document corpus takes 31 GB of RAM as float32 -- turbovec fits it in 4 GB and searches it faster than FAISS. The project is built on top of Google Research's TurboQuant, an algorithm presented at ICLR 2026 that achieves near-optimal compression with zero training data required.

The problem with vector search at scale

A collection of 10 million vectors stored in standard float32 format can easily require more than 30 GB of RAM. As datasets continue growing, memory costs, infrastructure requirements, and search latency become major bottlenecks. The standard solution has been Product Quantization (PQ), which compresses vectors by learning a codebook from the data. Existing vector quantization algorithms present a trade-off: either they lack accelerator compatibility and exhibit slow computation, making them unsuitable for real-time AI applications, or they suffer from suboptimal distortion bounds relative to bit-width.

Traditional vector quantization usually introduces its own memory overhead, as most methods require calculating and storing quantization constants for every small block of data. This overhead can add 1 or 2 extra bits per number, partially defeating the purpose of vector quantization. TurboQuant was built specifically to eliminate this overhead.

The math behind training-free compression

turbovec's real differentiator is not its compression ratio but the fact that it needs no training. FAISS product quantization has to learn a k-means codebook from data; TurboQuant randomly rotates each vector so its coordinates converge to a known distribution, then computes quantization boundaries from mathematics rather than from the data. This is what the paper calls a "data-oblivious" quantizer -- it works the same way on any dataset, without ever seeing your data first.