If you like Milvus but not running it, Zilliz Cloud is managed Milvus from the company behind the project. If you want to keep self-hosting with fewer moving parts, Qdrant and Weaviate run as single servers that can be clustered later. If the vectors belong to an application that already uses PostgreSQL, pgvector needs no new service at all. For local or in-application use, LanceDB and Chroma run inside your process. And if you want no infrastructure and accept a closed source service, Pinecone is serverless.
Milvus is an Apache 2.0 open source vector database hosted by the LF AI & Data Foundation; version 3.0.0 was released on 29 July 2026 and 3.0.5 on 30 October 2026. It has the widest documented choice of index types among open source vector databases (FLAT, the IVF family, HNSW with quantization, DISKANN, SCANN and GPU indexes), BM25 full-text and multi-vector hybrid search, and three deployment modes: the embedded Milvus Lite library, a single Standalone server, and a Distributed cluster on Kubernetes. The project itself sells nothing; the managed version is Zilliz Cloud.
The reasons people look elsewhere are mostly about operating it. The first two are documented by Milvus; the third is our editorial reading.
- Operational weight of distributed mode. A Distributed deployment depends on object storage, etcd and a write-ahead log service (Kafka, Pulsar or Woodpecker) as well as the Milvus components themselves, which needs Kubernetes and operations skill.
- More decisions. Choosing and tuning among a dozen index types is power that many retrieval projects do not need, and tenancy by partition or partition key does not support role-based access control, so strong per-tenant isolation means separate databases or collections.
- Scale that never arrived. Teams that chose Milvus for billions of vectors sometimes find their collections fit comfortably on one machine, where a simpler system costs less to run (editorial).
If your reason for leaving is a closed source managed service rather than operational weight, see our Pinecone alternatives page instead. For head-to-head pages, see Milvus vs Qdrant, Chroma vs LanceDB, Pinecone vs Weaviate and pgvector vs Pinecone.
Quick picks
Managed Milvus on AWS, Google Cloud and Azure, from a Free plan to Dedicated clusters and BYOC.
Apache 2.0 Rust server that runs as one Docker container and adds sharding and replication when needed.
No new service: vectors, HNSW indexes and SQL filters in your existing database.
Apache 2.0 library that stores tables in the Lance format on local disk or object storage, with no server.
Apache 2.0, runs in-process or as a single-node server, with a serverless Chroma Cloud for later.
How we chose
We considered Zilliz Cloud, Qdrant, Weaviate, Pinecone, pgvector, LanceDB, Chroma, Vespa, Elasticsearch, OpenSearch and FAISS. To be included, a product had to be actively developed or sold in November 2026, store vectors with metadata and serve similarity queries, and need less infrastructure to operate than a Distributed Milvus cluster: a managed service, a single server process, an embedded library or an extension to a database you already run.
Vespa, Elasticsearch and OpenSearch were left out because they are themselves multi-node clusters with their own sizing work, so they do not answer the operational complaint; they are covered on the Pinecone alternatives page and in Elasticsearch vs OpenSearch for vector search. FAISS was left out because it is a library without persistence, metadata storage or replication.
For each product we recorded the deployment modes, licence, pricing model, scaling approach and its main limitation. The order is not a ranking: entries start with managed Milvus, then single-server databases, then the PostgreSQL extension, then embedded libraries, then a fully managed proprietary service.
At a glance
| Product | How it runs | Licence | Scaling | Pricing model | Main trade-off |
|---|---|---|---|---|---|
| Zilliz Cloud | Managed service or BYOC | Commercial (engine Apache 2.0) | Serverless or CU-sized dedicated clusters | Free; $4 per million vCUs; Dedicated from $126 a month | AUTOINDEX only; some plans quote-only |
| Qdrant | Single server or cluster; Qdrant Cloud | Apache 2.0 | Sharding and replication | Free self-hosted; cloud free tier then hourly usage | Fewer index types than Milvus |
| Weaviate | Single server or cluster; Weaviate Cloud | BSD-3-Clause (wl/ folder commercial) | Sharding and replication; shard per tenant | Free self-hosted; cloud from $45 a month | In-memory HNSW needs RAM planning |
| pgvector | Extension inside PostgreSQL | PostgreSQL License | PostgreSQL instance and replicas; Citus for sharding | Free; cost is the PostgreSQL host | Index dimension limits; no built-in hybrid search |
| LanceDB | In-process library; Enterprise managed or BYOC | Apache 2.0 (Enterprise commercial) | Object storage in embedded mode; distributed in Enterprise | Free; Enterprise by quote | No self-serve cloud plan documented |
| Chroma | In-process or single-node server; Chroma Cloud | Apache 2.0 (Cloud paid) | Self-hosted single-node; Cloud distributed | Free; Cloud usage-based from $0 plus usage | Hybrid Search API is Cloud only |
| Pinecone | Proprietary serverless service; BYOC on Enterprise | Closed source | Serverless; dedicated read nodes on paid plans | Free Starter; $20 Builder; usage-based from $50 minimum | No self-hosted edition for most buyers |
Zilliz Cloud
Zilliz Cloud is the commercial managed service built on Milvus, from Zilliz. It offers a Free plan (one cluster, 5 GB storage, 2.5 million vCUs a month, 5 collections), Serverless clusters billed at $4 per million vCUs for reads and writes plus storage, Dedicated clusters from $126 a month on Standard and $197 a month on Enterprise, and Business Critical, On-demand Compute and BYOC deployments priced by sales.
It is the smallest change for an existing Milvus user: the data model and API stay Milvus, while etcd, object storage and the WAL service become the provider's problem (editorial). New work-email sign-ups get $100 of credits for 30 days.
- Vector index choice is reduced to AUTOINDEX (customised HNSW or DiskANN), so tuned self-hosted index settings do not carry over.
- Free clusters are suspended after 7 days of inactivity.
- If a trial ends without a payment method, clusters go to the recycle bin and are deleted after 30 days.
Qdrant
Qdrant is an Apache 2.0 vector database written in Rust. A single server process holds the data, a filterable HNSW index and typed payload indexes; sharding and replication are added when a collection outgrows one node. Hybrid queries combine dense and sparse prefetches with RRF or DBSF fusion, with BM25 computed server-side.
Compared with Milvus, the main difference is what you run: no separate metadata store or log service to operate (editorial). Qdrant Cloud adds a free-forever single-node cluster and an hourly usage-based Standard tier, and Hybrid Cloud and Private Cloud run Qdrant on your own infrastructure. See Milvus vs Qdrant.
- A single filterable HNSW index type, plus quantization, rather than Milvus's range of IVF, DiskANN and GPU indexes.
- Payload indexes should exist before HNSW is built; adding them later means rebuilding the index.
- Standard cloud unit rates appear only in the pricing calculator.
Weaviate
Weaviate is a Go vector database with most of its repository under BSD-3-Clause. It runs as a single Docker container or a Kubernetes deployment with sharding and replication, and offers HNSW, flat, dynamic and the disk-resident HFresh index, BM25 hybrid search and native multi-tenancy with one shard per tenant.
Two things set it apart for teams coming from Milvus: model provider modules that create embeddings inside the database, and per-tenant shards that give isolation without Milvus's choice between partitions (no RBAC) and separate collections (editorial). Weaviate Cloud starts with a Free plan of 100,000 objects and Flex from $45 a month.
- The default HNSW index is held in memory, so large deployments need RAM planning or compression.
- Files in the wl/ folder are under a commercial licence.
- Fewer documented index types than Milvus, which matters if you rely on DiskANN or GPU indexes.
pgvector
pgvector adds vector types, distance operators and HNSW and IVFFlat indexes to PostgreSQL. If the application that produces your embeddings already uses PostgreSQL, the vector store becomes a column and an index rather than a separate system, with SQL filtering, joins and transactions included.
It is the option with the least new infrastructure on this page, provided your data fits the limits of one PostgreSQL primary and its replicas; the README points to Citus for sharding and to partitioning or separate tables for tenants. See pgvector vs Pinecone.
- Indexes are limited to 2,000 dimensions for vector (4,000 for halfvec).
- Filtered approximate queries can return fewer rows unless iterative scans are enabled.
- No built-in BM25 hybrid search or embedding generation; combine PostgreSQL full-text search and your own code.
LanceDB
LanceDB is an Apache 2.0 embedded vector database: it runs inside your application, with no server, and stores tables in the Lance columnar format on local disk or object storage such as S3. It offers IVF and IVF-HNSW indexes with several quantization options, BTREE, BITMAP and LABEL_LIST scalar indexes, SQL filters, and BM25 full-text plus hybrid search fused with RRF.
Keeping data on object storage without a cluster makes it a practical replacement for Milvus in batch, analytics and single-application retrieval workloads (editorial). Note that the self-serve LanceDB Cloud is no longer documented: in November 2026 the docs list only LanceDB Embedded and LanceDB Enterprise (managed, BYOC or custom, by quote). See Chroma vs LanceDB.
- No published self-serve cloud plan or free tier; managed use means an Enterprise contract.
- Embedded mode has no built-in multi-user server; multitenancy features are not documented for the library.
- Some indexes are built only by the embedded library, not by Enterprise deployments.
Chroma
Chroma is an Apache 2.0 vector database that runs in-process, with local persistence, or as a single-node server, with an HNSW index, metadata and document filters, and built-in embedding functions (a local all-MiniLM-L6-v2 model by default). It is the simplest option on this page to start with (editorial).
Chroma Cloud is serverless and distributed, using a SPANN index; Starter is $0 a month plus usage with $5 of credits, Team is $250 a month plus usage, and usage is billed per GiB written, GiB-month stored and TiB queried. The Search API for hybrid retrieval is Cloud only as of November 2026.
- The documented self-hosted modes are single-node; the distributed architecture is offered only as Chroma Cloud.
- Hybrid search through the Search API is Cloud only for now.
- Fewer index choices than Milvus.
Pinecone
Pinecone is a proprietary serverless vector database: no servers, shards or index parameters to manage, with dense, sparse and BM25 full-text fields in document indexes (generally available September 2026) and hosted embedding and reranking models. Billing on Standard and Enterprise is by read units, write units and storage.
It is the opposite end of the scale from self-hosted Milvus: almost no operations work, but no choice of index algorithm and no self-hosted edition outside Enterprise BYOC. See Pinecone vs Weaviate.
- Closed source; leaving later means a migration rather than a redeploy.
- Read unit billing makes cost depend on query volume and result size.
- The free Starter plan is limited to AWS us-east-1, 5 indexes and 2 GB of storage.
When to stay with Milvus
Stay with Milvus if you genuinely need its range: billions of vectors, GPU indexes, DiskANN for collections larger than memory, or index-level tuning that other systems do not expose. If the problem is only the cluster, you may not need a new database: Milvus Standalone runs as a single server and Milvus Lite as an embedded library, and Zilliz Cloud runs the distributed version for you.
Switching makes most sense when your collections fit on one node and nobody wants to operate etcd, object storage and a log service (Qdrant, Weaviate), when the data belongs in PostgreSQL (pgvector), when retrieval runs inside one application (LanceDB, Chroma), or when you want no infrastructure at all (Pinecone). Plan to rebuild indexes on the new system, map partitions or partition keys to its tenancy model, and compare result quality on your own queries before cutting over.
Frequently asked questions
Is there a managed version of Milvus?
Yes. Zilliz Cloud, from Zilliz, is the commercial managed service built on Milvus, with Free, Serverless, Dedicated, Business Critical and BYOC options on AWS, Google Cloud and Azure. The Milvus project itself does not sell a managed service.
Can I run Milvus without Kubernetes?
Yes. Milvus Standalone runs as a single server (commonly in Docker) and Milvus Lite is an embedded Python library. Distributed mode is the one that needs Kubernetes, object storage, etcd and a WAL service.
Is Qdrant simpler to run than Milvus?
In our reading of the documentation, yes for most deployments: Qdrant runs as one server process and adds sharding and replication when needed, while Milvus Distributed depends on several external services. Milvus offers more index types in return. See Milvus vs Qdrant.
What is the lightest alternative to Milvus?
LanceDB and Chroma run inside your application with no server, and pgvector adds vectors to a PostgreSQL database you already have. FAISS is lighter still, but it is a library without persistence or metadata storage, not a database.
Sources
- Milvus repository
- Milvus index types
- Milvus release notes
- Milvus multi-tenancy
- Zilliz Cloud pricing
- Zilliz Cloud AUTOINDEX
- Qdrant indexing
- Qdrant pricing
- Weaviate vector indexes
- Weaviate pricing
- pgvector repository
- LanceDB documentation
- LanceDB Enterprise
- Chroma pricing
- Chroma Search API (Cloud)
- Pinecone pricing
Checked November 2026.
How we research these guides: our editorial method.