Milvus
Open-source distributed vector database for billion-scale search, from embedded Milvus Lite to Zilliz Cloud
Overview
Milvus is one of the most widely adopted open-source vector databases, built for very large scale. Its distributed architecture separates storage and compute so each layer scales on its own, and it supports a wide range of index types (HNSW, IVF variants, DiskANN for indexes larger than memory, and GPU indexes) along with metadata filtering, hybrid dense plus sparse search, full-text search and multi-vector queries.
Milvus was created by Zilliz, open-sourced in 2019, and is a graduated project of the LF AI & Data Foundation. It comes in several shapes: Milvus Lite (a Python library), Standalone (single node), Distributed (Kubernetes) and Zilliz Cloud, the fully managed service from its creators. Companies including Salesforce, Walmart, eBay and NVIDIA are listed as users.
The Verdict
Who Should Use Milvus?
Best For
- Hundreds of millions to billions of vectors
- Teams that want open source with no lock-in
- On-premises or bring-your-own-cloud requirements
- Teams with Kubernetes and platform experience
Not Ideal For
- Small projects that don't want to operate a cluster
- Teams without infrastructure experience (use Zilliz Cloud)
- Apps that need vectors joined with relational data
What's Great
- Proven at billion-vector scale
- Widest choice of index types, including GPU and DiskANN
- Runs anywhere, from a laptop to Kubernetes to managed cloud
- Hybrid dense, sparse and full-text search
- Large community and LF AI & Data governance
Watch Out For
- Distributed Milvus is complex to operate
- Steeper learning curve than fully managed services
- Several moving parts (object storage, message queue, etcd)
- Tuning indexes for recall and cost takes expertise
Pricing
View all features & details
Key Features
- HNSW, IVF, DiskANN and GPU indexes
- Dense, sparse and hybrid search
- Full-text search (BM25)
- Metadata filtering, JSON and array fields
- Multi-tenancy and RBAC
- Separate storage and compute scaling
Deployment Options
- Milvus Lite (pip install)
- Standalone (Docker)
- Distributed (Kubernetes)
- Zilliz Cloud managed service
How It Compares
| Feature | Milvus | Qdrant | Weaviate | Pinecone |
|---|---|---|---|---|
| Open source | Apache 2.0 | Apache 2.0 | BSD-3 | No |
| Billion-scale | Yes | Yes | Yes | Yes |
| GPU indexes | Yes | Partial | No | No |
| Managed option | Zilliz Cloud | Qdrant Cloud | Weaviate Cloud | Native |
| Operational effort | High (distributed) | Moderate | Moderate | Low |
Is this your tool? Add a badge