● Beta — KubeEra is in active development. Profiles are live; AI-perception measurement is rolling out.
KubeEra
M

Milvus

AI Agents
Open source · Apache-2.0

Milvus is an open-source vector database built for storing, indexing, and querying dense vector embeddings at scale.

updated 2026-07-01

What it is

Milvus is an open-source vector database built for storing, indexing, and querying dense vector embeddings at scale. It's the backing store for similarity search — the core primitive behind RAG pipelines, recommendation systems, image/audio search, and anomaly detection. It separates compute and storage, runs on Kubernetes natively, and supports multiple index types (HNSW, IVF, DiskANN, etc.) so you can tune the recall/latency/memory tradeoff per collection. Despite the "AI Agent" category tag here, it's infrastructure — a database, not an agent framework.

Who builds it and why

Originally developed by Zilliz, Milvus was open-sourced and later donated to the LF AI & Data Foundation. Zilliz still drives the majority of commits and also sells a managed cloud version (Zilliz Cloud), which is the standard commercial-open-source pattern: build the moat in the open, monetize the operational burden. CNCF maturity level is not publicly available in the data we have — check LF AI & Data's project listing directly if that matters for your procurement checklist, since Milvus lives under LF AI & Data, not CNCF proper.

Production readiness signal

45K GitHub stars and 359 contributors is a real, active community — not a side project. Last commit dated 2026-07-01 indicates continuous development (assuming clock accuracy in your source data). Latest release v2.6.19 suggests a mature versioning cadence well past initial 1.0 growing pains. Apache-2.0 license means no copyleft surprises for commercial use. Written in Go, which tracks with the broader cloud-native tooling ecosystem and makes it a reasonably natural fit if your platform team already runs Go-based infra.

What's not in this data: SLA history, benchmark independence, or CNCF graduation status. If you're evaluating for a regulated environment, get those answers directly from Zilliz or run your own load tests — don't take vendor benchmarks at face value.

Who should use this

  • Teams building RAG or semantic search at a scale where a single-node vector index (FAISS, Annoy, or pgvector on a small Postgres instance) won't cut it — think tens of millions to billions of vectors.
  • Platform teams that want a Kubernetes-native vector store they can operate themselves, with horizontal scaling of query nodes independent of storage.
  • Organizations already running multi-tenant AI infra who need collection-level isolation, RBAC, and dynamic schema support.
  • Anyone who needs index flexibility (swapping HNSW for IVF-PQ under memory pressure) without re-architecting.

Who should NOT use this

  • Small-scale projects with under a few million vectors — pgvector or SQLite with a vector extension will be less operational overhead and you won't need a distributed system.
  • Teams without dedicated infra capacity to run a distributed system with etcd, Pulsar/Kafka, and MinIO/S3 dependencies. Self-hosted Milvus is not a single binary — it's a distributed architecture with real operational weight.
  • Latency-critical, small-embedding-set use cases where an in-process library (FAISS embedded directly in your app) avoids network hops entirely.
  • Teams wanting a fully managed, zero-ops experience without touching Kubernetes — go straight to Zilliz Cloud or a competitor's managed offering instead of self-hosting.

Alternatives

  • Weaviate — similar cloud-native vector DB with built-in hybrid search and a more opinionated schema/module system.
  • Qdrant — Rust-based, simpler operational footprint, strong for teams wanting less infra complexity than Milvus's distributed architecture.
  • pgvector — Postgres extension; the right call if you already run Postgres and don't need billion-scale vector counts.

Pricing

Milvus itself is fully open source under Apache-2.0 — no license cost, no feature gating. You pay in infra and ops time to run it yourself (etcd, object storage, message queue, compute nodes). Zilliz Cloud offers a managed, usage-based hosted version if you want the API without the operational overhead — pricing for that is not publicly available in this dataset; check Zilliz's site directly.

Frequently asked

What is Milvus?+
Milvus is an open-source vector database built for storing, indexing, and querying dense vector embeddings at scale. It's the backing store for similarity search — the core primitive behind RAG pipelines, recommendation systems, image/audio search, and anomaly detection.
Who builds Milvus?+
Originally developed by Zilliz, Milvus was open-sourced and later donated to the LF AI & Data Foundation. Zilliz still drives the majority of commits and also sells a managed cloud version (Zilliz Cloud), which is the standard commercial-open-source pattern: build the moat in the open, monetize the operational burden.
Is Milvus production ready?+
45K GitHub stars and 359 contributors is a real, active community — not a side project. Last commit dated 2026-07-01 indicates continuous development (assuming clock accuracy in your source data). Latest release v2.6.19 suggests a mature versioning cadence well past initial 1.0 growing pains. Apache-2.
Who should use Milvus?+
Teams building RAG or semantic search at a scale where a single-node vector index (FAISS, Annoy, or pgvector on a small Postgres instance) won't cut it — think tens of millions to billions of vectors.
Who should not use Milvus?+
Small-scale projects with under a few million vectors — pgvector or SQLite with a vector extension will be less operational overhead and you won't need a distributed system. Teams without dedicated infra capacity to run a distributed system with etcd, Pulsar/Kafka, and MinIO/S3 dependencies.
What are the alternatives to Milvus?+
Weaviate — similar cloud-native vector DB with built-in hybrid search and a more opinionated schema/module system. Qdrant — Rust-based, simpler operational footprint, strong for teams wanting less infra complexity than Milvus's distributed architecture.
How much does Milvus cost?+
Milvus itself is fully open source under Apache-2.0 — no license cost, no feature gating. You pay in infra and ops time to run it yourself (etcd, object storage, message queue, compute nodes).