Best Vector Databases in 2026

Qdrant is an open-source vector database written in Rust, under Apache 2.0, that you can run yourself or rent as Qdrant Cloud. It filters on payload fields during the graph search, so a filtered query does not lose results the way post-filtering does, and it supports named vectors, sparse vectors and multi-vector (late-interaction) search in the same collection. Qdrant 1.19 in August 2026 added memory tiers for keeping less of a large index in RAM. Qdrant's own installation docs require block storage with a POSIX filesystem and rule out NFS and S3 for its data, so plan for local or block disks when you self-host.

Pinecone is a fully managed vector database: you create an index and write vectors, and there are no servers to size. Standard is billed by usage, with read units, write units and storage, over a $50 monthly minimum. In September 2026 Pinecone made full-text search generally available in the same database and launched a bring-your-own-cloud option that runs it in your own cloud account. It is closed source.

pgvector is an open-source PostgreSQL extension that adds a vector column type and HNSW and IVFFlat indexes. Your vectors sit in the same tables as the rest of your data, so a query can filter, join and search in one SQL statement, inside your existing backups and permissions. Version 0.8.0 added iterative index scans, which keep scanning when a filter removes too many results, and 0.8.7 shipped on October 1, 2026. Most managed Postgres services include it.

Milvus is an open-source distributed vector database, and Zilliz Cloud is its managed service. It separates storage from compute: index files and logs live in object storage (MinIO, S3 or Azure Blob), and query and data nodes scale apart. It offers the widest choice of index types here, including HNSW, IVF variants, DiskANN and GPU indexes. Milvus 2.6 reached GA on Zilliz Cloud in January 2026.

Weaviate is an open-source vector database with hybrid search built in: one query combines BM25 keyword scoring, vector similarity and filters. Its vectorizer modules can call an embedding model during ingestion, so you can write objects and let Weaviate create the vectors. Weaviate Cloud has a free tier and a Flex plan that starts at $45 a month, billed by vector dimensions and storage, and its September 2026 price book added RQ-4 compression on HNSW indexes.

turbopuffer is a managed search engine that stores vectors and documents on object storage and caches what is queried, so storage costs about as much as the bucket underneath. It runs vector, BM25 and filtered search. turbopuffer reports 200 ms p99 query latency over 100 billion vectors for its ANN v3 index (January 2026), a vendor figure we have not reproduced. Plans start at a $16 monthly minimum.

S3 Vectors is a vector index built into Amazon S3, generally available since December 2025. You create a vector bucket and indexes in it and query through the S3 API, paying for storage, uploads and queries with no cluster to run. An index holds up to 2 billion vectors, and a bucket up to 10,000 indexes. It is the cheapest place on this list to keep vectors, and it is built for corpora that are queried less often.

LanceDB is an open-source vector database built on the Lance columnar format. It can run inside your application with no server, against a local directory or an object store path (s3://, gs://, az://), and keeps vectors next to the raw data and metadata in one table. Lance commits rely on the storage's atomic put or rename so that exactly one concurrent writer wins. LanceDB sells a managed enterprise product, and its pricing page goes to a sales contact rather than a price list.

Chroma is an open-source vector database known for the shortest path to a working prototype: it runs in your Python process, persists to a local directory, and has a simple API. Chroma Cloud adds a hosted service with published usage prices, and in 2026 it added an EU region and Chroma Sync for pulling data from S3, GitHub and the web. Chroma's own docs describe the local persistent client as meant for development and testing, and recommend a server for production.

Vespa is an open-source search and serving engine that combines vector search, BM25 text search, structured filters and machine-learned ranking in a single query, and runs it at very large scale. It is the most capable engine here for ranking logic, and the hardest to learn: applications are defined in Vespa's own schema and configuration. Vespa Cloud prices by vCPU, memory and disk per hour.

Elasticsearch added dense vector fields and approximate kNN search to its text engine, so teams already running it can add semantic search without a second system. Elastic Cloud Serverless bills search by compute hour and storage per GB-month, and lists a vector-optimized profile that includes 50 GB free. It suits combining a mature text-search setup with vectors more than serving vectors alone.