Fastest embedded vector database in 2026: zvec, turbovec, USearch and FAISS compared
The fastest embedded vector database depends on the job: zvec is an in-process database; its benchmark shows 3.2x to 4.8x the throughput of Faiss IVF at matched recall. turbovec reports 4-bit search 4.4x faster than FAISS FastScan in less RAM. USearch reports 173,823 QPS on SIFT 100M. FAISS leads on GPUs. XERJ suits agents searching files.
Key number: On 2026-11-17 turbovec had 17,499 GitHub stars and zvec 16,261; turbovec reports 4-bit search 4.4x faster than FAISS IndexPQFastScan at matching recall. (as of 2026-11-17 · source: github.com/RyanCodrai/turbovec)
Last checked: 2026-11-17. Every third-party fact on this page was re-read from the tool's own docs or repository on that date.
TL;DR The fastest embedded vector database depends on your job. Choose zvec for a full database inside your app. Choose turbovec to fit a large corpus in little RAM. Choose USearch for huge graphs and many languages. Choose FAISS for GPUs and the widest index menu. If your agent mainly searches code, docs and data files, XERJ may be all you need.
Act as a coding agent. Read https://xerj.org/llms.txt and https://xerj.org/compare/fastest-embedded-vector-database-2026, then recommend zvec, turbovec, USearch, FAISS or XERJ for this project using the table on that page. If XERJ fits: install it, start a node with --embed-mode neural, index this repository with xerj autoindex, and run three searches. Report the recommendation and the reasons.
zvec (created December 2025) and turbovec (created March 2026) each passed 16,000 GitHub stars within a year, and Google suggests zvec vs faiss, turbovec vs faiss and usearch vs faiss. All four run inside your process with Python bindings. For server databases, see Weaviate, Qdrant and Milvus compared.
zvec, turbovec, USearch or FAISS
zvec
An in-process vector database from Alibaba with WAL storage, filters, full text search and dense plus sparse vectors. pip install zvec.
turbovec
A Rust index with Python bindings that stores vectors at 2 or 4 bits. Its README shows ten million 768-dimension embeddings in 4 GB.
USearch
A header-only C++ 20 proximity graph with bindings for Python, Rust, Go, JavaScript and more. ClickHouse and DuckDB embed it.
FAISS
Meta's C++ library with the widest index menu and GPU indexes that drop in for CPU ones.
XERJ
One binary with BM25 keyword search and kNN vectors, xerj autoindex for folders and xerj mcp for agents.
Side by side
| zvec | turbovec | USearch | FAISS | XERJ | |
|---|---|---|---|---|---|
| Kind | in-process database | quantized index | graph index | index library | search engine |
| Written in | C++ | Rust | C++ | C++ | Rust |
| Speed claim (own docs) | 3.2x to 4.8x Faiss IVF throughput at matched recall | 4-bit search 4.4x FAISS FastScan | 173,823 QPS on SIFT 100M | GPU indexes | BM25 and kNN in one node |
| Extras | WAL, filters, full text, sparse vectors | 4 GB for 10M 768-d vectors | RAM, VRAM and NVMe residency | widest index menu | autoindex, MCP, Elasticsearch compatible API |
| Licence | Apache-2.0 | MIT | Apache-2.0 | MIT | Apache-2.0 |
| Stars (2026-11-17) | 16,261 | 17,499 | 4,390 | 41,797 | 19,212 |
Speed claims come from each project's own benchmark, on its own hardware and data. They compare each tool with FAISS, so read them side by side with care.
When to choose zvec
Choose zvec when you want a real database inside your app. It keeps data safe with write-ahead logging and filters on structured fields. It also runs full text search over dense and sparse vectors. Its September 2026 benchmark shows 3.2x to 4.8x the throughput of Faiss IVF at matched Recall@10.
When to choose turbovec
Choose turbovec when RAM is the limit. It stores vectors at 2 or 4 bits. Its README shows ten million 768-dimension embeddings in 4 GB instead of 31 GB. Its 4-bit search runs 4.4x faster than FAISS FastScan at matching recall. It compares the query with every vector, so pair it with a graph index for hundreds of millions of vectors.
When to choose USearch
Choose USearch for very large graphs or many languages. One header-only C++ 20 core serves Python, Rust, Go, JavaScript, Java and more. One checkpoint opens in RAM, VRAM or on NVMe. ClickHouse, DuckDB, ScyllaDB and YugabyteDB embed it.
When to choose FAISS
Choose FAISS for GPU search, research work and the widest set of index types. It is the reference library the other three measure against, with 41,797 GitHub stars.
XERJ: a new way for AI to search for data
XERJ is a search engine for agents. One static binary holds documents with BM25 full-text search and kNN vector search over HNSW. It has an Elasticsearch compatible API. Its default embedder is lexical feature hashing, and --embed-mode neural loads all-MiniLM-L6-v2 on the node. XERJ is single-node.
xerj autoindex indexes a folder of code, docs and data, and xerj mcp gives agents search tools. For raw vector speed at scale, the four libraries above are the better fit. More: XERJ vs a vector database.
How this page was checked
Speed claims come from each project's README and benchmark docs, read on 2026-11-17: the zvec RaBitQ post, the turbovec README and the USearch README. Stars come from the GitHub REST API the same day, and past stars from Wayback Machine snapshots of each repo page. We did not run a vector benchmark for this page.
Try it with XERJ
Use the Copy the prompt button and paste it into your coding agent. It recommends one of the five for your project and, if XERJ fits, indexes your repository and runs three searches.
Read next
FAISS, ChromaDB and LanceDB compared, Qdrant, pgvector and ChromaDB compared and local RAG setup.
FAQ
What is the fastest embedded vector database in 2026?
Each project wins its own benchmark. zvec reports 3.2x to 4.8x the throughput of Faiss IVF at matched recall, turbovec reports 4-bit search 4.4x faster than FAISS FastScan, and USearch reports 173,823 QPS on SIFT 100M. Test on your own vectors and recall target.
zvec vs FAISS: which should I use?
Choose zvec for a database inside your app with storage, filters and full text search. Choose FAISS for a pure index library with GPU support and the widest set of index types.
turbovec vs FAISS: what is the difference?
turbovec quantizes every vector to 2 or 4 bits and scans them with SIMD, so it suits large corpora in little RAM. FAISS offers graph, IVF and GPU indexes for very large collections with strict latency.
Which one runs in Rust?
turbovec is written in Rust and installs with cargo add turbovec. zvec has a Rust crate, zvec-rust, and USearch has Rust bindings.
Do I need a vector database for RAG?
For agents that search code, docs and data files, a search engine with BM25 and kNN vectors often covers it. Read vector database vs full text search for the tradeoffs.
How was this page checked?
Speed claims come from each project's README and benchmark docs, read on 2026-11-17. Stars come from the GitHub REST API the same day, and past stars from Wayback Machine snapshots of the repo pages. We did not run a vector benchmark for this page.
Evidence
- zvec is an open source, in-process vector database from Alibaba with dense and sparse vectors, full text search, structured filters, write-ahead logging and pip install zvec. Source: https://github.com/alibaba/zvec
- In zvec's own benchmark with 8 threads at matched Recall@10, 7-bit IVF-RaBitQ delivers 3.2x to 3.7x the throughput of Faiss on GIST-1M and 3.8x to 4.8x on Cohere-1M. Source: https://zvec.org/en/blog/2026-09-08-zvec-rabitq/
- turbovec is a vector index written in Rust with Python bindings. Its README says ten million 768-dimensional embeddings take 31 GB as float32 and turbovec holds them in 4 GB, and 4-bit search runs 4.4x faster than FAISS IndexPQFastScan at matching recall. Source: https://github.com/RyanCodrai/turbovec
- turbovec scans every vector; its README suggests an HNSW or IVF index in front of it for hundreds of millions of vectors behind strict latency. Source: https://github.com/RyanCodrai/turbovec
- USearch is a header-only C++ 20 proximity graph index. On 192 AWS Graviton5 cores with SIFT 100M it builds at 161,952 vectors/s and searches at 173,823 QPS, while FAISS HNSW builds at 84,536 and searches at 178,600. ClickHouse, DuckDB, ScyllaDB and YugabyteDB embed it. Source: https://github.com/unum-cloud/USearch
- FAISS is a library for efficient similarity search and clustering of dense vectors, with GPU indexes that can replace CPU indexes. Source: https://github.com/facebookresearch/faiss
- GitHub stars on 2026-11-17 from the GitHub REST API: FAISS 41,797; turbovec 17,499; zvec 16,261; USearch 4,390. Wayback Machine snapshots of the repo pages show turbovec at 15,859 on 2026-08-20 and FAISS at 40,738 on 2026-08-14. Source: https://api.github.com/repos/RyanCodrai/turbovec
Related
- FAISS vs ChromaDB vs LanceDB in 2026: local vector search compared
- Qdrant vs pgvector vs ChromaDB in 2026: which vector store for a RAG app?
- Should my agent call a search API or stand up Qdrant for persistent memory?
- Local RAG: answer questions over your own files, on your machine
- Weaviate vs Qdrant vs Milvus in 2026: open source vector databases compared · same intent: vector database