Overview
Milvus is the CNCF open-source vector database, and it takes the route of solving scale first and convenience second. Its distributed architecture separates compute, storage and coordination, so each layer scales independently and the growth path stays legible once a cluster gets large.
Free self-hosting, independent scaling and the ability to index directly on lakehouse formats make it comfortable for larger teams that already run a data platform.
Key Features
- High scale ceiling: Designed for billion-scale vector retrieval with GPU indexing and multi-replica support, and one of the earlier projects in the category to address scale explicitly.
- Decoupled distributed architecture: Compute, storage and coordination are separate layers that scale independently, so capacity planning and cost control can be handled separately.
- Managed option on Zilliz Cloud: The free tier includes 5GB of storage, 2.5 million vCU per month and 5 collections; Standard starts at $0 for serverless or $126/month for dedicated; Enterprise starts at $197/month with a 99.95% SLA, SAML SSO, RBAC and VPC peering.
- Billed by compute unit: Three tiers — performance-optimised at $63, capacity-optimised at $16 and tiered storage at $5 — priced per million 768-dimension vectors per month, giving real freedom to pick by scenario.
- Indexing on lakehouse formats: Version 3.0, released July 29, 2026, introduced lake-native capability that indexes directly on open table formats including Iceberg, Lance, Parquet and Vortex, removing a data movement step.
- Community and adoption: Over 45,700 GitHub stars, with users including Salesforce, Reddit, Walmart and DoorDash.
Use Cases
- Billion-scale retrieval that needs a distributed cluster rather than a single-node service
- Teams with a data lake that want to index on Iceberg or Lance instead of copying data again
- Self-hosting to control cost while retaining a legible horizontal scaling path
- Enterprise retrieval requiring SSO, RBAC and VPC peering
Pros
- Apache 2.0 open source with no licence cost to self-host
- Designed for billion-scale retrieval with a clear scaling path
- Indexes on open table formats such as Iceberg and Lance since 3.0
- Managed free tier available, with paid tiers split by compute unit
Pricing
Free to self-host under Apache 2.0. Zilliz Cloud free tier: 5GB storage, 2.5M vCU per month, 5 collections. Standard: serverless from $0, dedicated from $126/month. Enterprise from $197/month. Compute units: performance-optimised $63, capacity-optimised $16, tiered storage $5, per million 768-dimension vectors per month.
Summary
Milvus targets teams that have already reached scale. Free self-hosting, independent scaling and lakehouse indexing make it natural in environments that already run a data platform. Conversely, for a light application with a few thousand vectors the architecture is heavy, and the managed entry price sits above peers with free tiers — that scenario is better served starting on Chroma or Qdrant's free tier.
Version History
- Milvus 3.0 (2026-07-29): Introduced lake-native capability, indexing directly on Iceberg, Lance, Parquet and Vortex
- Zilliz Cloud tiers (2026): Free tier at 5GB plus 2.5M vCU per month; Enterprise from $197/month