Milvus
Multi-vector columns (10 simultaneous). Native hybrid search (v2.5). CAGRA + Vamana GPU/CPU (v2.6).
At a glance
- Type
- Multi-vector + hybrid + GPU
- Tier
- T1
- Section
- Vector-database infrastructure
- Created
- 2019-09
- Latest release
- v2.6.15 2026-04-17
- License
- Apache-2.0
- GitHub
- 44.1k★ +546/mo Go
- Pricing
- Free (OSS Apache-2.0) + Zilliz Cloud (pay-per-use managed)
- Funding
- $113M total (Zilliz: $43M 2020 + $60M 2022); Zilliz is the commercial entity
Taxonomy
- storage
- vector
- retrieval
- similarity
- persistence
- long-term
- update
- overwrite
- unit
- chunk
- governance
- inspectable
- conflict
- overwrite
When to use
Optimised for: low-latency similarity search + scale
Anti-fit: not for relational / graph-heavy queries (vector-first by design)
Pros & cons
Pros
Largest-scale OSS vector DB — billions of vectors comfortably; CNCF graduate; strong community in China + research labs.
Cons
Operational complexity is high vs Pinecone-managed; documentation can be uneven across language clients.
Claims & capabilities
40,000+ GitHub★ milestone reinforcing leadership in OSS vector DBs; 10,000+ enterprise deployments (Alibaba, Cisco, BIGO/Likee, Rakuten Symphony); Milvus 2.5 (2025) added native hybrid search; production reports sub-50ms retrieval at billion-vector scale; Zilliz Cloud benchmarks claim 10x faster ANN vs Weaviate at 5B+ vectors with sub-10ms latency
Technical surface
- API surface
- REST, gRPC, SDK: Python, Java, Go, Node.js, C#, Ruby
- Backend storage
- custom (object-storage backed; etcd metadata; Pulsar/Kafka log)
- Deployment
- Both (self-hosted OSS + Zilliz Cloud managed)
- Embedding model
- multiple supported
- Multi-tenancy
- Control plane and data plane in separate VPCs/subnets; BYOC option for hard tenant isolation
- MCP
- via official adapter — milvus-mcp
- A2A
- no Google A2A (Agent2Agent) integration documented as of 2026-05.
- OpenTelemetry
- first-class — Prometheus + OTel
Similar systems
Other vector-database infrastructure in the catalog, ranked by inbound references.
- Qdrant T1
Distribution-Based Score Fusion + RRF. Sparse vectors native; filtering via ANN graph modification.
- pgvector T1
Stores embeddings alongside relational + full-text data. HNSW + IVFFlat ANN indexes. Used as agent conversation memory via LangChain + MCP. Foundation of Supabase AI and many self-hosted RAG stacks.
- Pinecone T1
Managed vector DB. Cascading sparse + dense + rerank pipeline; pinecone-rerank-v0 .
- Chroma T1
Limited native hybrid (users build RRF custom). Fast Rust core (v2.5).
- LanceDB T1
Embedded vector DB (Arrow columnar). RRF reranker. Petabyte-scale on disk.
- MongoDB Atlas Vector Search T1
Agent memory store for both short-term (document) and long-term (vector). LangGraph checkpointer for stateful agents. Vector search extended to Community Edition (Sept 2025).
Related systems
Referenced by (1)
- NVIDIA ReMEmbR builds on — storing captions with timestamps + 3D position coordinates in MilvusDB