Turbovec, an Open-Source Rust Implementation of Google's TurboQuant, Claims to Outperform FAISS in Vector Search Benchmarks
An open-source Rust vector index built on Google Research's TurboQuant algorithm reports beating FAISS on speed, memory and insertion latency in benchmarks published alongside its 1.0.0 release.
Overview
An independent developer has released turbovec, an open-source Rust vector index that implements TurboQuant, a quantization algorithm developed by Google Research, for vector search rather than the large-language-model memory compression it was originally built for. The project’s documentation states that a 10-million-document corpus that takes 31 GB of RAM as float32 fits in 4 GB under turbovec, and that it “searches it faster than FAISS,” the library the project uses as its benchmark baseline throughout its published results. The GitHub repository lists the project as MIT-licensed with more than 15,550 stars, and its package manifest shows it has reached version 1.0.0. The MIT license file credits the project to developer Ryan Codrai.
What We Know
Turbovec is described in its own documentation as “a Rust vector index with Python bindings, built on Google Research’s TurboQuant algorithm — a data-oblivious quantizer with near-optimal distortion and no separate training phase,” according to the project’s README. It is distributed via pip install turbovec for Python and cargo add turbovec for Rust, and it ships drop-in replacements for the in-memory vector or document stores used by four popular retrieval-augmented-generation (RAG) frameworks: LangChain’s InMemoryVectorStore, LlamaIndex’s SimpleVectorStore, Haystack’s InMemoryDocumentStore, and Agno’s LanceDb integration, per the README.
On compression, the README states that a 1,536-dimension vector shrinks from 6,144 bytes in 32-bit floating point to 384 bytes at 2-bit quantization, a 16x reduction, the project reports. On search speed, turbovec’s own benchmark suite — run on Google Axion ARM chips and Intel Sapphire Rapids x86 chips — found it beating a comparably configured FAISS IndexPQFastScan index in every tested configuration, averaging “3.4× at 4-bit and 23% at 2-bit across the eight cells of each width, on both architectures,” according to the same documentation. Broken out by architecture, the project reports averaging 3.5x faster at 4-bit and 26% faster at 2-bit on ARM, and 3.4x faster at 4-bit and 20% faster at 2-bit on x86, per the README.
The same benchmarks report that inserting a single vector into a populated index takes 6.3 to 19.7 microseconds depending on configuration, which the project describes as 7.6 to 13.9 times faster than a single insert into FAISS, according to the documentation. Removing a vector by ID is reported at 0.44 to 1.37 microseconds per operation in turbovec, against 0.19 to 1.02 seconds for the equivalent FAISS operation at 100,000 vectors, the README states, a gap the project attributes to FAISS repacking its stored codes on every removal call. On search-result quality, the documentation reports that its calibrated quantization variant beats FAISS on recall-at-1 in three of four tested cells using OpenAI-sized embeddings, while trailing by 0.7 points in the fourth, per the same source.
The algorithm itself traces to a paper titled “TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate”, authored by Amir Zandieh, Majid Daliri, Majid Hadian and Vahab Mirrokni and first posted to arXiv on April 28, 2025, according to the paper’s own listing. The paper states that its quantizer achieves distortion rates “differing only by a small constant ($\approx 2.7$) factor” from the information-theoretic lower bound, and that in nearest-neighbor search tasks its method “outperforms existing product quantization techniques in recall while reducing indexing time to virtually zero,” per the paper’s abstract. Turbovec’s README lists the paper’s presentation venue as ICLR 2026.
Google Research’s original presentation of TurboQuant focused on a different application: compressing the key-value cache used during large-language-model inference, rather than vector-search indexing. As previously reported by The Machine Herald, that March announcement described up to a 6x reduction in KV-cache memory and triggered a sell-off in memory-chip stocks, but noted there was “no confirmed timeline for integration into Google’s own products or cloud services” and that independent developers had already begun building their own implementations for other use cases. Turbovec applies the same underlying quantization approach to a different problem: indexing and searching dense vector embeddings for retrieval-augmented generation, rather than compressing model memory during inference.
What We Don’t Know
Turbovec has not tagged a formal release on GitHub; its version number is tracked only in its Rust package manifest, which currently reads 1.0.0. All of the performance figures described above come from benchmarks the project ran and published itself, using its own methodology and hardware selection; they have not been independently reproduced or verified by a third party. It is not yet clear whether any vector database vendor or RAG framework maintainer plans to adopt turbovec as a default backend rather than an optional integration, or how the library performs on datasets and query patterns outside the GloVe and OpenAI-embedding benchmarks the project has published so far.