Vector Database Comparison: 7 Self-Hosted Tools—What Actually Differs

All seven vector databases in this comparison run offline, support self-hosting, and carry permissive open-source licenses. But they diverge on hardware requirements, deployment architecture, and maturity. This guide walks through the real differences so you can pick based on your actual constraints.

Entry-Level: Start With 2GB RAM (Chroma, Qdrant, LanceDB)

If you're prototyping a RAG pipeline or embedding store on a single machine, Chroma, Qdrant, and LanceDB all run on 2GB RAM with no GPU required.

Chroma is Python-native and embeds directly into your application. It's the gentlest ramp for Python developers who want vector search without a separate service. Mature, actively maintained.

Qdrant is written in Rust and runs as a standalone service with a REST and gRPC API. Faster than Chroma at scale, very active development, same 2GB floor. Pick this if you want to decouple the vector store from your application language.

LanceDB is the newer entry—also embedded, Apache-licensed, serverless-oriented. If you're building with DuckDB or want columnar analytics alongside vector search, LanceDB fits that niche. Active maintenance.

Mid-Tier: 4GB+ RAM, More Features (Weaviate, Vespa)

Weaviate requires 4GB+ RAM and adds GraphQL as its query interface. Modular architecture lets you swap components. Good if your team is already comfortable with GraphQL and wants a schema-driven vector store. Mature, steady development.

Vespa also needs 4GB+ RAM but positions itself as a hybrid search engine—vector search plus traditional full-text and field filters in one engine. Scales well, active development. Use Vespa if you're handling queries that mix vector similarity with keyword filters or ranking rules.

Distributed & Kubernetes-Scale (Milvus, Vald)

Milvus recommends 8GB+ RAM and is designed for distributed deployments. Written in Go, k8s-friendly, handles massive collections. Actively maintained by the community. Pick this if you're already running Kubernetes and need to scale horizontally.

Vald is cloud-native from the ground up—architecture assumes Kubernetes, uses a distributed ANN search model. Best for teams running container infrastructure at scale. Less common in small teams, but very active if you commit to k8s.

Decision Checklist

• Single machine, Python app: Chroma
• Single machine, multi-language app: Qdrant
• GraphQL interface required: Weaviate
• Hybrid vector + keyword search: Vespa
• Kubernetes cluster: Milvus or Vald
• Embedded, serverless-style: LanceDB

License and Maintenance Signal

All seven carry Apache-2.0 or BSD-3-Clause licenses and support offline operation. Qdrant and Milvus show the highest commit velocity; Chroma, Weaviate, and Vespa are mature and stable. Vald and LanceDB are active but younger. Last verification: 2026-09-12. Check the source repository (linked in each tool's row) for the latest commit date and issue triage before final selection.

Originally published at Forged Goods. The ready-made version: Local-AI Stack Directory: 40 Self-Hosted LLM & Vector-DB Tools, Verified Specs.