Milvus vs Qdrant: Which Open-Source Vector Database Should You Run?

Milvus and Qdrant are both free, open-source, Apache 2.0-licensed vector databases you can self-host with no licensing cost — the same starting point as most vector-database comparisons, and not where this one gets decided.

Milvus is a distributed, cloud-native vector database built for scalable ANN (approximate nearest neighbor) search at very large scale (milvus-io/milvus, GitHub). It was originally built by Zilliz in 2017, then contributed to the Linux Foundation's LF AI & Data Foundation in 2020, where it graduated as a top-level project in 2021. Qdrant is a vector database and search engine written in Rust, built by the company of the same name, known for fast, filtered search that blends metadata filters with vector similarity at query time.

Both offer a managed cloud on top of the open-source core — Zilliz Cloud for Milvus, Qdrant Cloud for Qdrant — so the choice isn't really open source versus managed. It's distributed-at-massive-scale versus fast, filter-heavy retrieval in a simpler footprint.

Milvus vs. Qdrant: Side-by-Side

License and governance: both open source, different ownership models

Neither database charges to self-host. Milvus and Qdrant are both released under the Apache 2.0 license, which permits commercial use, modification, and self-hosting at any scale with no licensing fee.

The governance model differs. Milvus was contributed to the Linux Foundation's LF AI & Data Foundation in 2020 and graduated as a top-level project in 2021 — the same neutral-foundation model used by projects like Kubernetes. Qdrant remains a company-led open-source project, developed and maintained primarily by Qdrant itself rather than a foundation.

That distinction matters mostly for procurement teams weighing long-term project continuity, not for day-to-day usage — both projects are active, and Apache 2.0 means your right to self-host doesn't depend on either company's survival.

Architecture: distributed scale versus fast filtering

Milvus is described by its own maintainers as a high-performance, cloud-native vector database built for scalable ANN search (milvus-io/milvus, GitHub). It's designed from the ground up as a distributed system — you can run it as a cluster spread across nodes, which is what lets it target billions of vectors.

Qdrant is a single Rust binary at its core, which makes it simpler to run for small and mid-size workloads. Its standout feature is fast, filtered search: it applies metadata filters at query time without a heavy performance penalty, which matters for RAG applications that need to narrow results by tenant, date, or category before ranking by similarity.

In practice, this means Milvus tends to win on raw scale and Qdrant tends to win on query flexibility per vector at a given scale. Neither claim requires taking either vendor's benchmark at face value — test both against your own data and filter patterns before committing.

Pricing and hosting: what each managed cloud actually offers

Self-hosted, both are free. The managed clouds are where the offerings diverge in structure, not necessarily in cost.

Qdrant Cloud publishes a free-forever tier limited to a single node (0.5 vCPU, 1GB RAM, 4GB disk) — enough for prototyping and small projects, not production scale. Above that, the Standard tier is usage-based with a 99.5% uptime SLA, and Premium adds SSO and private VPC links with a 99.9% SLA and a minimum spend (Qdrant). Qdrant Hybrid Cloud lets you run a Qdrant-managed cluster on your own infrastructure, and Private Cloud is fully custom.

Zilliz Cloud (the managed Milvus offering) publishes a permanent free tier of its own: 5GB of storage, 2.5 million vCUs (Zilliz's usage-credit unit) per month, and up to 5 collections (Zilliz). Above that, it moves to usage-based serverless pricing plus dedicated clusters for guaranteed capacity at production scale.

• Qdrant Cloud free tier: single node, 0.5 vCPU / 1GB RAM / 4GB disk, free forever (Qdrant)
• Qdrant Cloud paid tiers: Standard (usage-based, 99.5% SLA), Premium (min spend, SSO, private VPC, 99.9% SLA)
• Zilliz Cloud (Milvus) free tier: 5GB storage, 2.5 million vCUs/month, up to 5 collections, permanent (Zilliz)
• Both: self-hosting the open-source core is always free, at any scale

→ Compare per-token pricing for every AI model

Which one should you actually run?

Pick Milvus when your vector count is large enough — tens or hundreds of millions of vectors and up — that you need a distributed architecture built for that scale from day one, and you're comfortable either running a cluster yourself or paying Zilliz Cloud to run it for you.

Pick Qdrant when your workload is small to mid-size, your queries lean heavily on metadata filtering (multi-tenant apps, date-bounded search, category-scoped RAG), or you want the simplest possible self-hosted footprint to start with — a single binary instead of a distributed system.

If you're not sure which bucket you're in yet, that's itself useful information: Qdrant's free-forever tier and single-binary setup make it the cheaper, faster way to prototype. You can always migrate to Milvus later if your scale genuinely outgrows it.

The Verdict

Milvus and Qdrant both remove the vendor-lock-in question — Apache 2.0, self-hostable, no licensing fee either way. The real decision is architectural: Milvus is a distributed system built for very large-scale vector search, while Qdrant is a faster, simpler system built around filtered retrieval.

Most teams starting a new RAG project are not yet at Milvus's target scale, and Qdrant's free-forever tier and single-binary setup make it the cheaper place to start. Teams that already know they're heading toward hundreds of millions of vectors should evaluate Milvus (or Zilliz Cloud) directly against that requirement rather than retrofitting later. Not sure which fits your actual data volume and RAG architecture? Layer3 Labs can help you size it before you commit.