Milvus
activeby LF AI & Data Foundation (developed mainly by Zilliz) · since 2019 · also known as Milvus Lite
Cloud-native, distributed open-source vector database built for billions of vectors.
Overview
Milvus is the most widely deployed open-source vector database for large scale, graduated at the LF AI & Data Foundation and developed mainly by Zilliz. Its disaggregated architecture separates streaming, query and data nodes over object storage, and it supports many index types (HNSW, IVF, DiskANN, GPU CAGRA, sparse) plus BM25 full-text and hybrid search. Milvus 3.0 (July 2026) added online schema changes, external collections over Lance/Vortex data, faceted search and new sparse index algorithms. Milvus Lite runs embedded in Python for prototyping; the distributed version is serious infrastructure that benefits from Kubernetes expertise.
Strengths & trade-offs
Strengths
- +Proven at billion-vector scale with independent scaling of query and ingest
- +Widest choice of index types including DiskANN and GPU (CAGRA)
- +BM25 full-text and sparse vectors for hybrid search
- +Tunable consistency levels per query
- +Milvus Lite gives the same API embedded in Python
Trade-offs
- −Distributed mode has many moving parts (etcd, object storage, message log)
- −Heavy for small projects compared with Qdrant or Chroma
- −Strict collection schemas need planning despite new online changes
- −Major-version upgrades (2.x to 3.x) need care with storage formats
Great for
- +Billion-scale semantic and image search
- +RAG platforms serving many teams
- +Recommendation and deduplication
- +Multimodal retrieval over lakehouse data
Think twice for
- −Transactional or relational workloads
- −Single-developer side projects (use Milvus Lite or Chroma)
- −Teams without Kubernetes experience self-hosting at scale
Capabilities
Native or official first-party support · 11 of 25 tracked
Scale & consistency
- Comfortable data size
- Massive · Multi-PB
- Horizontal scaling
- Native
- Multi-region
- No
- Transactions
- None
- Consistency
- Tunable, Strong, Eventual
- Schema
- Strict
- Architecture
- Distributed, Shared storage, Embedded, Client server
- Query languages
- Milvus API (boolean filter expressions)
Compute nodes scale independently over object storage; deployments with tens of billions of vectors are documented.
Deployment
available on · AWS · Google Cloud · Azure
Ecosystem
Alternatives to Milvus
Fully managed Milvus from its creators, with serverless, dedicated and BYOC deployments.
milvus vs zilliz-cloud →Rust vector search engine with rich payload filtering, quantization and a managed cloud.
milvus vs qdrant →Open-source AI-native vector database with hybrid search, modules for embeddings and multi-tenancy.
milvus vs weaviate →Fully managed serverless vector database with built-in embedding, reranking and assistants.
milvus vs pinecone →Big-data serving engine for search, recommendation and RAG with tensors and ML ranking.
milvus vs vespa →Lucene-based distributed search and analytics engine for text, vectors, logs and metrics.
milvus vs elasticsearch →Milvus FAQ
What is Milvus?
Milvus is a vector database by LF AI & Data Foundation (developed mainly by Zilliz), first released in 2019. Milvus is the most widely deployed open-source vector database for large scale, graduated at the LF AI & Data Foundation and developed mainly by Zilliz. Its disaggregated architecture separates streaming, query and data nodes over object storage, and it supports many index types (HNSW, IVF, DiskANN, GPU CAGRA, sparse) plus BM25 full-text and hybrid search. Milvus 3.0 (July 2026) added online schema changes, external collections over Lance/Vortex data, faceted search and new sparse index algorithms. Milvus Lite runs embedded in Python for prototyping; the distributed version is serious infrastructure that benefits from Kubernetes expertise.
Is Milvus open source?
Yes. Milvus is open source. License: Apache License 2.0.
Does Milvus support ACID transactions?
No. Milvus does not provide ACID transactions; its consistency model is tunable, strong, eventual.
Does Milvus scale horizontally?
Milvus scales horizontally out of the box, with built-in sharding or partitioning. It is comfortable at massive scale (Multi-PB).
How can I run Milvus?
You can self-host it, embed it in your application and use the vendor's managed cloud. Managed offerings include Zilliz Cloud. It is available on AWS, Google Cloud and Azure.
What is Milvus best for?
Milvus is a strong fit for billion-scale semantic and image search; RAG platforms serving many teams; and recommendation and deduplication. Think twice for transactional or relational workloads.
What are the best alternatives to Milvus?
Common alternatives to Milvus include Zilliz Cloud, Qdrant, Weaviate, Pinecone and Vespa. The right choice depends on your workload, scale and hosting; WhichDB's smart match ranks them against your requirements.
How popular is Milvus?
Milvus ranks #35 of 844 databases on the WhichDB Score (45.9/100), and #2 among vector databases. The score combines public adoption signals such as package downloads, GitHub activity, job posts and developer discussion.