Milvus

Milvus is an open-source vector database built for high-performance similarity search at scale, from millions of vectors on a single machine to billions of vectors spread across a distributed cluster. Created by Zilliz and donated to the LF AI & Data Foundation in 2020, it is distributed under Apache 2.0 and developed by a full-time engineering team with contributions from over 200 open-source developers worldwide. More than 10,000 enterprise organizations run Milvus in production, including NVIDIA, Salesforce, eBay, Airbnb, and DoorDash. As of April 2026, the current stable release is Milvus 2.6.x.

The core capability is approximate nearest neighbor (ANN) search over high-dimensional vector embeddings, with support for 11+ index types including HNSW, IVF_FLAT, DiskANN, SCANN, and GPU-accelerated CAGRA via NVIDIA's cuVS library. Milvus 2.5 (December 2024) added native full-text search using Sparse-BM25, enabling hybrid vector-keyword queries in a single engine without a separate Elasticsearch deployment. Milvus 2.6 (June 2025) introduced Woodpecker, a cloud-native write-ahead log that eliminates the Kafka/Pulsar external dependency, plus hot/cold tiered storage for up to 50% storage cost reduction. Zilliz also operates Zilliz Cloud, a fully managed Milvus service with a free tier, serverless pay-as-you-go, and dedicated clusters starting at $99/month.

What Milvus actually does in April 2026

Milvus runs vector similarity search across datasets that break other tools. The project has been through two major architectural rebuilds: the original monolithic 1.x line and the fully disaggregated cloud-native 2.x architecture that ships today. The 2.x design separates storage, compute, and coordination into independent layers, so you can scale query nodes separately from data ingestion nodes based on actual workload shape.

Index support is the widest in open-source vector databases. Beyond HNSW, you get IVF variants for memory-constrained deployments, DiskANN for on-disk billion-scale indexes, SCANN for high-recall scenarios, and four GPU-accelerated index types through NVIDIA cuVS. The GPU CAGRA index delivers roughly 10x higher throughput on batch search workloads compared to CPU-only HNSW at comparable recall levels. For teams with NVIDIA GPU infrastructure, this is not a marginal improvement.

Milvus 2.5's full-text search integration is the most significant architectural addition in recent releases. Prior to 2.5, a common production pattern was running Milvus alongside Elasticsearch - Milvus for semantic search, Elasticsearch for keyword matching - and merging results in the application layer. Milvus 2.5 replaced this with built-in Sparse-BM25 powered by Tantivy indexing. The benchmark Zilliz published at launch showed Milvus 2.5 returning results in 6 milliseconds versus Elasticsearch at 200 milliseconds on 1 million vectors, a 30x improvement. Teams that upgraded eliminated an entire cluster from their infrastructure.

Milvus 2.6 tackled the infrastructure dependency problem from a different angle. The new Woodpecker WAL system eliminates external Kafka or Pulsar requirements by persisting log data directly to object storage (S3, GCS, MinIO). Woodpecker achieves 450 MB/s throughput in local filesystem mode, 3.5x faster than Kafka, and 750 MB/s against S3, 5.8x faster than Kafka. This simplification matters most for teams self-hosting who previously had to operate and monitor a Kafka cluster in addition to etcd and object storage.

"Teams depend on Milvus in demanding environments where performance, reliability, and cost all matter. As adoption grows, we'll keep listening closely to the community and turning that trust into a platform that scales for enterprise production." - James Luan, VP of Engineering at Zilliz, December 2025

Where Milvus sits versus Qdrant and Pinecone

These three cover the main decision space for production vector databases, and they differ mechanically in ways that matter for deployment choices.

Qdrant is written in Rust and ships as a single binary with zero external dependencies for standalone deployment. Its core ANN index is HNSW only - one algorithm, well-tuned, with rich metadata filtering built on top. Benchmarks on 384-dimensional SQuAD vectors show Qdrant query latency at 94ms vs. Milvus at 250ms (Qdrant wins on per-query latency), but Milvus achieves approximately 46 QPS vs. Qdrant at 4.7 QPS (Milvus wins on throughput by roughly 10x). Qdrant is the better choice for sub-100M vector workloads where single-digit millisecond latency matters more than aggregate throughput. Milvus pulls ahead when your ingestion volumes, query concurrency, or total vector count require horizontal scaling - Qdrant's horizontal clustering is simpler than Milvus's but not designed for the same data volumes. Qdrant also has no GPU acceleration support. Milvus supports NVIDIA CAGRA for 10x throughput gains on compatible hardware.

Pinecone is fully proprietary and closed-source, with no self-hosted option. You use it through their managed cloud service only. Pinecone abstracts the indexing layer entirely - you cannot choose index type, tune HNSW parameters, or access storage internals. For teams with no infrastructure mandate, this is a feature: Pinecone reaches production in an afternoon. For teams that need data portability, audit access, or recall-latency tuning for specific workloads, the abstraction becomes a ceiling. Pricing diverges significantly at scale: Pinecone's managed service runs $700-1,200/month for 10M vectors, while equivalent Milvus self-hosted infrastructure typically runs $500-2,000/month depending on instance selection, with the difference being operational control vs. zero-ops convenience. 2026 benchmarks show Milvus at 8ms p95 latency with roughly 4,200 QPS versus Pinecone at 12ms p95 with roughly 2,800 QPS for comparable query workloads, giving Milvus a modest performance edge at scale when infrastructure is well-tuned.

"We're not just combining two search approaches - we're revolutionizing enterprise search with a solution that's 30 times faster while dramatically simplifying infrastructure." - Charles Xie, Founder and CEO of Zilliz, December 17, 2024

What the deployment reality looks like

Milvus ships in two modes. Standalone mode runs on Docker and suits development and small-scale workloads. Cluster mode is Kubernetes-native and is the recommended path for production. The gap between these two is significant.

A production Milvus cluster requires, at minimum: a Kubernetes cluster, etcd for metadata storage, and object storage (S3 or MinIO). Before Milvus 2.6, you also needed Kafka or Pulsar for the write-ahead log - that dependency is now replaced by Woodpecker, which is a meaningful simplification. The Milvus Operator (a Kubernetes operator maintained by Zilliz) handles much of the deployment scaffolding and lifecycle management. Using Helm directly without the operator is possible but fragile: default values are not production-ready, and tuning resource limits, replica counts, and storage classes before a real workload hits requires experience with both Kubernetes and distributed databases.

For teams without that expertise, Zilliz Cloud removes the operational burden entirely. The managed service offers the same Milvus API surface with a free tier (5 GB, 2.5M vCUs/month), a Serverless tier at $4 per million vCUs, and Dedicated clusters from $99/month. The Zilliz Cloud platform claims 99.95% uptime SLA and runs on AWS, Azure, and GCP. Organizations migrating from OpenSearch to Zilliz Cloud have reported up to 8x cost reductions while maintaining or improving search performance, according to Zilliz's June 2025 release announcement. The managed path is the realistic choice for teams that want Milvus's capabilities without the infrastructure investment.

Who Milvus is built for

Milvus is the natural choice when your vector search workload is genuinely large. Teams building RAG systems that need to search across hundreds of millions of document chunks, recommendation engines indexing hundreds of millions of products or media items, or multimodal search pipelines combining image and text embeddings at scale - these are the scenarios Milvus was engineered for. The GPU-accelerated indexes specifically benefit teams running NVIDIA hardware in their inference stack who also need the highest possible search throughput.

Data engineering teams with Kubernetes experience and a preference for open-source infrastructure control over managed convenience will find Milvus's architecture familiar and the flexibility valuable. The Apache 2.0 license and LF AI & Data Foundation governance mean no vendor lock-in on the data layer: your embeddings are yours, portable across deployments.

The growing integration ecosystem reinforces this: Milvus works natively with LangChain, LlamaIndex, Haystack, and LangGraph for agent and RAG pipelines, and connects to data platforms through Apache Spark and Kafka connectors for high-volume ingestion workflows.

What Milvus is not

Milvus is not the right choice for teams at early stage who need vector search running in an afternoon without Kubernetes expertise. Chroma (embedded Python, zero infrastructure) or Qdrant (single binary, simple REST API) get you to a working prototype faster. Milvus's architecture overhead only becomes an advantage when the scale, throughput, or index flexibility requirements justify it.

It is not a relational database with vector capabilities bolted on. Teams running primarily structured queries with vector search as a secondary filter may find pgvector (Postgres extension) more ergonomic - it keeps all data in one system and avoids a separate database to operate. Milvus shines when vector retrieval is the primary operation, not a secondary filter on SQL results.

Teams that have no intention of running their own infrastructure and do not need the open-source portability guarantees should evaluate Pinecone or Zilliz Cloud directly. There is no reason to absorb Milvus's operational complexity if the only benefit you need is a managed vector search endpoint with simple authentication and auto-scaling.