Vector databases store embeddings and return the nearest matches for retrieval and recommendations.
This guide ranks twelve of them on what the first ninety days involve: how much memory your vectors need, what the free tier actually holds, whether the engine can leave the vendor's cloud, and whether you need a separate database at all or can use Postgres.
Vendors can pay for visibility on this page. It never changes what an entry
says about a product, including the criticism, and we earn nothing when you click through to a
vendor. How that works.
In short
What vector databases software does
A vector database stores numerical embeddings of text, images or other data and finds the closest matches to a query vector quickly, usually alongside metadata filters.
In this order: setup effort, what it really costs, how your data comes back out, whether
you can leave, and who each vector databases tool is built for. One rule sits on top of those tests:
Places 1 to 3 go to challengers, vendors outside the best-known brands in a category, wherever the category has them; the established names follow from place 4 in their ranked order. Why those five, and why there is no
score out of ten, is on the how we work page.
Organisations that need keyword, filter and vector search in a single engine
Small teams who only want a simple vector index
Country is where the vendor is headquartered or contracts from, which is a
different question from where your data is hosted. Where the two tell different stories, the
entry says so.
Qdrant is built around one job and says so: the repository describes a vector database and vector search engine, licensed Apache-2.0. The managed cloud runs the same software on AWS, Azure or GCP, and Hybrid Cloud puts managed clusters on your own compute.
That makes exit cheap. The cost is scope: you still need another store for relational data, and the free tier is a prototype size, so budget for the paid tier early.
What stands out
Open source
Hybrid cloud
Free cluster
Where it costs you
The free cluster is limited to a single node with 1 GB of RAM and 4 GB of disk
Standard tier bills by usage, so the monthly figure comes from a calculator rather than a price list
Right for
Teams that want a dedicated vector engine they can move between cloud and own servers
Wrong for
Teams that want one database for text, joins and vectors
GermanyFree cloud cluster; usage-based paid tiers; free to self-host
Open-source vector database with hosted clusters and built-in embedding calls
Ranked #2 of 12 in Best Vector Databases in 2026.
Free tierOpen sourcePublished pricingEurope
Weaviate sells more than storage. The cloud bundles an embeddings service priced per million tokens and a query agent with its own monthly allowance, alongside the core database.
That shortens a first build, and it adds meters that sit next to the cluster fee. The code is BSD-3-Clause and runs locally in Docker, so the database itself stays portable even when the extras are not.
What stands out
Open source
Hosted embeddings
Free cluster
Where it costs you
The free cluster holds 100,000 objects, 1 collection and up to 3 tenants
Dedicated deployments are sold as prepaid contracts through sales, so scale pricing is not published
Right for
Teams that want embeddings, hybrid search and storage handled in one hosted service
Wrong for
Buyers who need dedicated-cluster pricing before talking to sales
NetherlandsFree sandbox cluster; Flex pay-as-you-go published; Premium by contract
Apache-licensed vector store that starts as a Python package
Ranked #3 of 12 in Best Vector Databases in 2026.
Free tierOpen sourceSelf-hostablePublished pricingNorth America
Chroma separates the library from the service. The Apache-2.0 repository installs with pip and offers client-server mode, while Chroma Cloud adds serverless search priced per GiB written, per GiB stored and per TiB queried.
That split gives a clean exit path. The usage model suits spiky traffic and punishes heavy write churn, so model your ingest volume before assuming the low entry price holds.
What stands out
Open source
Usage-based
Python-first
Where it costs you
The starter plan is capped at 10 databases and 10 team members
Single-tenant and bring-your-own-cloud clusters sit on the Enterprise tier
Right for
Developers who want to prototype locally and move to a hosted database later
Wrong for
Teams planning billions of vectors with strict uptime commitments
United StatesFree starter plan with credits; usage-based cloud; open source to self-host
Serverless vector and text search built on object storage
Ranked #4 of 12 in Best Vector Databases in 2026.
Published pricingElsewhere
The architecture is the pitch: data lives in object storage, hot data is cached in memory and on SSD, and you pay for what you query. That suits multi-tenant products with thousands of rarely touched indexes.
The trade is commitment and control. Minimums start small but rise by tier, there is no self-hosted option, and single-tenancy and bring-your-own-cloud are reserved for the top plan.
What stands out
Object storage
Cloud only
Minimum commitment
Where it costs you
Hosted by the vendor only; no downloadable edition exists to run yourself
Every plan carries a monthly minimum, and the enterprise tier adds a 35% usage premium
Right for
Applications with many namespaces where most data is idle most of the time
Wrong for
Buyers who need a self-hosted or open-source engine
Not stated by the vendorUsage-based with a monthly minimum per plan, published
Fully managed vector database with a free starter plan
Ranked #5 of 12 in Best Vector Databases in 2026.
Free tierPublished pricingNorth America
Pinecone sells operational relief. The Starter plan is free, Builder is a flat monthly fee, and Standard moves to a minimum monthly usage with a trial credit, then pay as you go for database, inference and assistant usage.
Security features such as SAML sign-on and backup and restore arrive at Standard, while customer-managed keys and bring-your-own-cloud wait for Enterprise. The price of that comfort is a single vendor with no engine you can take elsewhere.
What stands out
Fully managed
Free tier
Proprietary
Where it costs you
Proprietary service with no self-hosted edition on offer
Single sign-on, backups and read nodes are tied to the paid Standard plan
Right for
Teams that want a managed index with no servers or tuning to run
Wrong for
Buyers who need an open-source engine to run themselves
United StatesFree Starter plan; flat and minimum-usage paid plans, published
Milvus is the choice when volume is the problem. The repository describes separated compute and storage, horizontal scaling and several index types including HNSW and DiskANN, plus GPU indexing.
Smaller footprints are possible through Standalone mode or Milvus Lite. The catch is ownership: Zilliz is the major contributor, so roadmap pull comes from the company that sells the hosted version.
What stands out
Open source
Distributed
Self-hosted
Where it costs you
The distributed deployment is Kubernetes-native, which means real operational work
Managed hosting is a separate commercial product from the project, Zilliz Cloud
Right for
Platform teams running very large collections on their own Kubernetes
Wrong for
Small teams without anyone to run Kubernetes services
Not stated by the vendorFree to self-host; managed version sold as Zilliz Cloud
Managed Milvus with serverless, dedicated and bring-your-own-cloud options
Ranked #7 of 12 in Best Vector Databases in 2026.
Free tierSelf-hostablePublished pricingNorth America
Zilliz Cloud turns an engine that is hard to run into a service you rent, with Free, Standard, Enterprise and Business Critical plans and a 30-day trial on the paid tiers.
Dedicated and serverless clusters are priced in different units, which blurs the real bill. Moving back to self-hosted Milvus is possible because the engine is shared, but you inherit the operations work the service was hiding.
What stands out
Managed Milvus
Free plan
BYOC
Where it costs you
The free plan holds 5 GB of storage and up to 5 collections
Audit logs, SAML sign-on and private endpoints require the Enterprise plan
Right for
Teams that want Milvus without running Kubernetes themselves
Wrong for
Anyone who needs one flat monthly price to budget against
United StatesFree plan; serverless and dedicated rates published; enterprise tiers
In-memory search engine with vector search and auto-embedding
Ranked #8 of 12 in Best Vector Databases in 2026.
Free tierOpen sourceSelf-hostablePublished pricingNorth America
Typesense appeals when search quality for typed queries matters as much as nearest-neighbour recall. The documentation covers importing your own embeddings, plus auto-embedding with built-in models, OpenAI, Azure OpenAI or other compatible APIs.
Cloud pricing follows the memory you provision, with no limit on records or operations. The in-memory design is quick, and it means collection size maps straight to RAM cost.
What stands out
Open source
In-memory
Typo-tolerant
Where it costs you
The index is held in memory, so large collections need large and costly clusters
GPL-3.0 licensing may constrain embedding the server inside distributed products
Right for
Product search where typed queries and vectors must work together
Wrong for
Archives of hundreds of millions of vectors that must stay cheap
United StatesFree to self-host; cloud clusters sized by memory; free cluster option
pgvector removes a system from the stack. The README lists exact and approximate nearest-neighbour search, HNSW and IVFFlat indexes, half-precision, binary and sparse vectors, and the Postgres features around them.
Exit is trivial, because the data is already in Postgres. The limit is that vector load competes with your transactional workload, and nobody sells you tuning help unless your Postgres host does.
What stands out
Postgres extension
Open source
Self-hosted
Where it costs you
No vendor, hosting or support contract of its own comes with the extension
Exact search is the default, so large tables need an HNSW or IVFFlat index to stay fast
Right for
Teams already on Postgres whose vectors belong beside their relational data
Wrong for
Teams that want a managed service with a support contract
Not stated by the vendorFree extension; cost is whatever your Postgres hosting charges
Hosted Postgres platform with vector columns through pgvector
Ranked #10 of 12 in Best Vector Databases in 2026.
Free tierPublished pricingAsia-Pacific
Supabase makes sense when vectors are one feature of an application, not the application. The pricing page shows a free plan, a Pro plan from a flat monthly fee per project, and usage charges for egress and storage beyond the included amounts.
The code is Apache-2.0 on GitHub. The compromise is shared resources: heavy similarity queries run on the same database that serves everything else.
What stands out
Hosted Postgres
Free plan
pgvector
Where it costs you
Free projects pause after a week of inactivity and the plan allows two active projects
Vector search shares compute with the rest of the Postgres database
Right for
Product teams that want auth, storage and vectors from one Postgres backend
Wrong for
Workloads that need a dedicated high-throughput vector engine
SingaporeFree plan; per-project monthly plans plus usage, published
In-memory data platform with vector search over fields in Redis
Ranked #11 of 12 in Best Vector Databases in 2026.
Free tierPublished pricingNorth America
For an application already on Redis, vector search over existing fields avoids adding another service. The docs describe vector fields and search inside Redis, and the pricing page notes a Flex option that holds data at 10% RAM to cut cost.
The weak point is economics at scale: memory is the main price driver, and the licence for new code is a choice among three terms, only one of which is AGPL.
What stands out
In-memory
Free tier
Source available
Where it costs you
The free database is limited to 30 MB on a shared deployment with community support
Redis 8 code is offered under RSALv2, SSPL or AGPLv3, which restricts some commercial uses
Right for
Teams already running Redis who want vector queries beside their cache
Wrong for
Buyers who need cheap storage for very large vector collections
United StatesFree 30 MB cloud database; hourly or monthly paid plans, published
Search and analytics engine that also works as a vector database
Ranked #12 of 12 in Best Vector Databases in 2026.
Published pricingEurope
Elasticsearch fits when full-text relevance is already a requirement. The company's own page positions it as a vector database, and the serverless price list separates ingest, search, machine-learning compute, storage and egress, with a 50 GB allowance for vector profiles.
The repository is triple-licensed, with part under the Elastic License 2.0 alone. Expect a bigger learning curve and a harder cost forecast than the single-purpose engines above.
What stands out
Hybrid search
Serverless
Self-managed
Where it costs you
Serverless is billed on several meters at once: ingest, search, storage and egress
Some code in the x-pack folder carries only the Elastic License 2.0
Right for
Organisations that need keyword, filter and vector search in a single engine
A vector database stores numerical embeddings of text, images or other data and finds the closest matches to a query vector quickly, usually alongside metadata filters. The differences that matter are rarely in the feature list, so this is
the order we would work through them.
01
Decide whether you need a published price
12 of the 12 tools here publish what they cost; the other 0 quote per organisation. The ones you can compare without a sales call: Qdrant, Weaviate, Chroma, turbopuffer, Pinecone, Milvus, Zilliz Cloud, Typesense, pgvector, Supabase, Redis, Elasticsearch.
02
Decide how much the jurisdiction matters
These 12 vendors are established in 5 countries across 4 regions (North America 5, Europe 3, Elsewhere 3, Asia-Pacific 1). That decides whose disclosure law applies to what the vendor holds, wherever the servers are.
03
Consider whether you want the source
6 of these are open source: Qdrant, Weaviate, Chroma, Milvus, Typesense, pgvector. Hosting one yourself trades a subscription for maintenance.
Memory, not records, sets the vector database bill
Vectors are large, and where they live decides the invoice. Redis and Typesense keep data in RAM, which is quick and means collection size converts directly to memory cost; Redis lists a Flex option holding data at 10% RAM for that reason. Typesense Cloud sizes clusters by memory, not by record count.
turbopuffer goes the other way, keeping data in object storage with a memory and SSD cache, so idle indexes cost little. Qdrant's free cluster holds 1 GB of RAM, which shows how quickly a prototype outgrows free. Before comparing quotes, multiply your vector count by dimensions and bytes, then ask each vendor which tier that fits.
Work out total vector size from count, dimensions and numeric precision.
Ask whether the quote covers RAM, disk or object storage.
Check whether replicas multiply the bill.
Free tiers hold different things
A free tier proves the integration, not the workload, and each one counts something different. Weaviate caps at 100,000 objects, 1 GB of memory and one collection. Zilliz Cloud gives 5 GB of storage and five collections. Redis offers 30 MB. Chroma starts at no monthly fee but spends $5 of credits before usage billing begins.
Supabase pauses free projects after a week of inactivity. Pinecone's Starter plan is free, and Standard comes with a three-week trial credit. None of those units line up, so a catalogue of 200,000 products fits one and fails another. Load your real data during the trial instead of the sample set.
Load a realistic slice of your own data into each free tier.
Note which limit is hit first: objects, memory, storage or inactivity.
Find out what the upgrade step costs before you build on the free plan.
Can the engine leave the vendor's cloud
Exit in this market depends on the licence. Qdrant is Apache-2.0, Weaviate is BSD-3-Clause, Milvus is Apache-2.0 and Chroma is Apache-2.0, so you can run the code yourself if the hosted price changes. Typesense is GPL-3.0. Redis 8 and Elasticsearch are offered under several licences, each with a source-available option next to AGPL.
Pinecone and turbopuffer are hosted services without a downloadable edition, so leaving means exporting vectors and rebuilding the index elsewhere. That is not a flaw, but it is a switching cost, and it belongs in the comparison next to price. Ask each hosted vendor how a full export works and in what format.
Read the licence file in the repository, not the marketing page.
Test a full export from the hosted service during the trial.
Confirm what a re-index of your data costs in time and compute.
Do you need a vector database at all
For many applications the answer is no. pgvector adds vector columns, HNSW and IVFFlat indexes to Postgres, and Supabase offers the same inside a hosted backend. If your data is already in Postgres, that removes a system, a sync job and a second bill. Elasticsearch and Redis make the same argument for teams already running them.
A dedicated engine such as Qdrant, Weaviate, Pinecone or Milvus earns its place when vector volume, filtering speed or isolation between tenants outgrows the host database. The decision depends on load: similarity queries compete for the same server resources as your transactions, so test under realistic concurrency before choosing the one-database route.
Run similarity queries under your normal write load and watch latency.
List what you must keep in sync if vectors live in a separate system.
Decide who is on call for the vector service.
What goes wrong most often when buying vector databases software
Sizing a cluster from the number of documents instead of vectors times dimensions, then discovering at launch that memory is four times what was budgeted.
Building a prototype on a free cluster and finding the next tier is a different architecture, with different limits and a different bill.
Choosing a hosted-only service without testing a full export, and learning the format only when you want to leave.
Adding a dedicated vector database to a Postgres application that had a few hundred thousand vectors and no performance problem.
07
Frequently asked questions
8 answers
What is the best vector databases in 2026?
Qdrant leads our ranking of 12. A company that publishes its engine under Apache-2.0 and sells the same engine as a managed service, so leaving the cloud means pointing your clients at your own server.
The free cluster is small, and the production tier is usage-based with a calculator rather than a fixed price. Hybrid Cloud runs managed clusters in your own infrastructure. The engine does vector search and little else, so keyword-heavy search needs a second system.
Which vector databases tools publish their pricing?
12 of the 12, with the pricing model each one publishes:
Qdrant: Free cloud cluster; usage-based paid tiers; free to self-host.
Weaviate: Free sandbox cluster; Flex pay-as-you-go published; Premium by contract.
Chroma: Free starter plan with credits; usage-based cloud; open source to self-host.
turbopuffer: Usage-based with a monthly minimum per plan, published.
Pinecone: Free Starter plan; flat and minimum-usage paid plans, published.
Milvus: Free to self-host; managed version sold as Zilliz Cloud.
Which vector databases tools can you host yourself?
Qdrant, Chroma, Milvus, Zilliz Cloud, Typesense, pgvector. The other 6 are hosted by the vendor only.
What should you use instead of Qdrant?
Weaviate and Chroma are the next two on this page. Weaviate is for teams that want embeddings, hybrid search and storage handled in one hosted service; Chroma is for Developers who want to prototype locally and move to a hosted database later.
Who should not buy Qdrant?
Teams that want one database for text, joins and vectors. The free cluster is limited to a single node with 1 GB of RAM and 4 GB of disk.
If your vector databases product belongs among these 12, tell us what it does and who it is for. Inclusion is an editorial call; what a listing is and is not is set out under software advice.
