Buying snapshot
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Decision point | What the project states |
|---|---|
License and governance | Apache License 2.0. The project is part of the LF AI & Data Foundation, with Zilliz as its major contributor. |
Self-managed modes | Milvus Lite, a library that runs in notebooks or on laptops; Milvus Standalone on one machine; and Milvus Distributed as a cluster that scales out. |
Managed option | Zilliz Cloud runs Milvus as a fully managed service, with Serverless, Dedicated and bring-your-own-cloud (BYOC) options. The project says Zilliz Cloud can be tried for free. |
Installation | pip install -U pymilvus installs the Python SDK. Installing pymilvus[milvus-lite] adds Milvus Lite, which keeps data in a local file. |
Building from source | Linux (Ubuntu 20.04 or later recommended) or macOS on x86_64 or Apple Silicon. Needs Go 1.21 or later, CMake 3.26.4 or later but below 4, GCC 11 or later on Linux or llvm 15 or later on macOS, and Python above 3.8 up to 3.11. |
Security controls | User authentication, TLS encryption for network traffic and role-based access control (RBAC). |
Our take
Choose Milvus when vector search is central to your AI application and you want an open-source database you can host yourself, with Zilliz Cloud as a managed route if you later stop wanting to run it. It suits teams that need to combine dense, sparse and full text retrieval and want to pick their own index type.
Evaluate Milvus
Popular comparisons
See all 13 Milvus comparisons
Where it stands out
One client from laptop to production. MilvusClient opens a local file for Milvus Lite, or takes an endpoint and credentials for your own Milvus server or Zilliz Cloud. The website promises code that you write once and deploy to production with one line.
Several kinds of search in one collection. Dense and sparse vectors can sit in the same collection. Milvus supports full text search with BM25, learned sparse embeddings such as SPLADE and BGE-M3, metadata filtering, range search and multi-vector search, and you can define functions to rerank results from several search requests.
Index choice for different workloads. Supported index types include HNSW, IVF, FLAT (brute force), SCANN and DiskANN, with quantization-based variants and mmap. Milvus also supports GPU indexing, such as NVIDIA's CAGRA.
Scale reads and writes separately. Compute and storage are separate, so you add query nodes for read-heavy traffic and data nodes for write-heavy traffic. Replicas add fault tolerance and throughput. Frequently used data can stay in memory or on SSDs while less-used data moves to slower, cost-effective storage, and tenants can be isolated by database, collection, partition or partition key.
Tools around the database. Milvus integrates with LangChain, LlamaIndex, OpenAI and HuggingFace. Related tools include Attu for GUI administration, Birdwatcher for debugging, Prometheus and Grafana for monitoring, Milvus CDC for data synchronization, VTS for migration, and connectors for Spark, Kafka, Fivetran and Airbyte.
Product screenshot
Trade-offs to weigh
Self-hosted scale means running Kubernetes. Milvus Distributed is built as stateless microservices on Kubernetes. To go past one machine without Zilliz Cloud, your team operates that cluster, including its query nodes, data nodes and replicas.
Lite is for prototyping, not production. The project positions Milvus Lite for learning and prototyping, Standalone for production or testing with datasets of up to millions of vectors, and Distributed for billions of vectors. Estimate your dataset size first, because it decides whether one machine is enough.
Public signals
About these signals
Verified factual signals from public sources. They indicate observable activity or interest, not total adoption, product quality, or cost.
736 GitHub commits 90d46.6k GitHub stars0 vulnerabilities across 2 packages
See all signals from 8 sources
npm · @zilliz/milvus2-sdk-node@3.0.6 · PyPI · pymilvus@3.0.2
Alternatives to Milvus
The reviewed substitutes for Milvus among the vector databases, and what would make each one the better answer.
Direct alternatives
Reviewed substitutes: products bought for the same job, where a team picks one.
- LanceDB
- Milvus and LanceDB are open-source vector databases for search over embeddings in AI applications. Milvus is a distributed, cloud-native database with stateless components that scales storage and compute separately, while LanceDB keeps multimodal data with native versioning in the Lance columnar format on disk or object storage.Applies to: Choosing a vector store for embedding search in a retrieval or agent application.
- Pinecone
- Pinecone and Milvus are vector databases built for production similarity search at large scale. Pinecone is commercial and fully managed, so teams use it as an API without running infrastructure, while Milvus is open source and cloud-native and can be self-hosted on Kubernetes or bought as a managed service.Applies to: Choosing a vector store for embedding search in a retrieval or agent application.
- Weaviate
- Weaviate and Milvus are open-source vector databases that scale to very large collections and come with managed cloud offerings. Weaviate stores objects next to their vectors and combines keyword and vector search in one query, while Milvus separates storage and compute in a cloud-native design aimed at very high vector counts.Applies to: Choosing a vector store for embedding search in a retrieval or agent application.
- Qdrant
- Milvus and Qdrant are open-source vector databases that teams self-host or buy as a managed cloud service for production similarity search. Milvus is designed as a distributed system with storage and compute separated for very large collections, while Qdrant is written in Rust and is known for fast search with detailed payload filtering.Applies to: Choosing a vector store for embedding search in a retrieval or agent application.
- Turbopuffer
- Large-scale vector search for AI applications is the core job of both Milvus and Turbopuffer. Milvus is an open-source, cloud-native database that teams can run themselves, while Turbopuffer is a commercial serverless service that adds full-text search and keeps data on object storage behind a cache to hold down cost.Applies to: Choosing a vector store for embedding search in a retrieval or agent application.
Other approaches
A different approach to the same problem. Each substitutes only for the workload named beside it.
- Marqo
- Marqo optimises search conversion using click-stream, purchase and event data, creating a personalised experience that knows what your customers are looking for - better than they do.Applies to: For teams that build their own product search on a vector store rather than buying a hosted ecommerce search platform.
- pgvector
- Milvus and pgvector both store embeddings and run similarity search. pgvector adds vector search to a PostgreSQL database a team already runs, while Milvus is a dedicated, distributed vector database, so the question is whether the workload's scale justifies running another system.Applies to: Whether embedding search can reuse an existing datastore or needs a dedicated vector service.
Head to head
See detailed alternatives analysis
Looking for Milvus alternatives? Whether you need a different deployment model, tighter integration with existing infrastructure, or a managed service that reduces operational overhead, the vector database landscape offers several compelling options. Milvus is a well-regarded open-source vector database built for GenAI applications, featuring a cloud-native architecture with separated storage and computation. However, depending on your use case, team size, and infrastructure preferences, one of the alternatives below may be a stronger fit.
Top Alternatives Overview
Pinecone is a fully managed, purpose-built vector database designed for production-scale similarity search. It offers a usage-based pricing model with a free tier, removing the need to manage infrastructure. Pinecone focuses on delivering relevant results at any scale and is positioned as a turnkey solution for teams that want to avoid operational complexity.
Qdrant is an open-source vector search engine written in Rust, offering both self-hosted and cloud deployment options. It provides a hybrid cloud model and enterprise-grade features including RAG support, recommendation systems, and advanced search capabilities. Qdrant has accumulated over 33,000 GitHub stars, reflecting strong community adoption.
Weaviate is an open-source vector database that stores both data objects and vector embeddings from ML models. It supports billions of data objects and combines multiple search techniques, including keyword-based and vector search. Weaviate offers a managed cloud service alongside its open-source self-hosted option, with a focus on reducing hallucination and data leakage in AI-native applications.
pgvector is an open-source PostgreSQL extension that adds vector similarity search directly to Postgres. It supports exact and approximate nearest neighbor search, multiple distance metrics (L2, cosine, inner product), and both HNSW and IVFFlat indexing. With over 20,000 GitHub stars, pgvector is ideal for teams already running PostgreSQL who want to avoid introducing a separate database system.
LanceDB is an open-source, multimodal vector database built on the Lance columnar data format. It features persistent storage, zero-copy versioning, and native support for text, images, and other multimodal data types. LanceDB supports compute-storage separation and is designed for AI workloads ranging from search to model training.
Vespa is an open-source AI search platform designed for large-scale RAG, personalization, and recommendation workloads. It provides native tensor support for complex ranking and decisioning, with real-time inference capabilities at enterprise scale.
Zilliz offers a fully managed cloud version of Milvus, built by the creators of the Milvus project. It simplifies deployment and scaling by eliminating the need to maintain vector search infrastructure, and is available as both SaaS and BYOC (bring your own cloud) options.
Architecture and Approach Comparison
The architectural differences between these vector databases reflect fundamentally different philosophies about how vector search should be deployed and managed.
Milvus uses a distributed, cloud-native architecture where storage and computation are separated. All components are designed to be stateless, which enhances elasticity and horizontal scaling. This makes Milvus well-suited for large-scale deployments but introduces operational complexity for smaller teams.
Embedded vs. standalone vs. distributed is the first major architectural divide. pgvector takes the embedded approach, running as a PostgreSQL extension within your existing database. This means vector search shares the same ACID guarantees, backup systems, and operational tooling you already use for relational data. LanceDB similarly runs in-process and can be used as a library, with persistent storage backed by object storage like S3. Turbopuffer also builds on object storage, using a serverless architecture with memory and SSD caching layers for performance.
Pinecone and Zilliz Cloud represent the fully managed end of the spectrum. Both abstract away all infrastructure concerns, but Zilliz Cloud is notable as the managed version of Milvus itself, meaning migration between self-hosted Milvus and Zilliz Cloud is designed to be straightforward.
Qdrant and Weaviate occupy a middle ground, offering both open-source self-hosted deployments and managed cloud services. Qdrant is written in Rust, which contributes to its memory efficiency and performance characteristics. Weaviate combines vector search with structured filtering and supports multiple search techniques in a single query.
Vespa differentiates itself by combining vector search with real-time ML model inference and complex tensor operations, making it suitable for applications that need more than pure similarity search. Marqo is a different kind of product: a hosted ecommerce search platform that trains a model on a store's catalog and shopper behavior, rather than a database for embeddings the team creates.
For indexing strategies, most solutions support HNSW (Hierarchical Navigable Small World) graphs. pgvector offers both HNSW and IVFFlat, giving users a choice between faster queries (HNSW) and lower memory usage (IVFFlat). Milvus supports IVF, HNSW, and DiskANN, providing flexibility across different hardware configurations.
Pricing Comparison
Vector database pricing models vary significantly depending on whether you choose self-hosted open-source, managed cloud, or serverless options.
Open-source / self-hosted (no license cost): Milvus, pgvector, Qdrant, Weaviate, LanceDB, Vespa, and Typesense all offer open-source editions that can be self-hosted at no software licensing cost. The primary expense is your own infrastructure and operational overhead.
Managed cloud services introduce subscription or usage-based pricing. Pinecone uses a usage-based model with a free tier included. Weaviate Cloud offers a freemium model with a free 14-day sandbox, and its managed service starts at $45 per month. Zilliz Cloud provides a free tier and a Standard plan from $0 per month on Serverless, with its Enterprise tier from $197 per month (Dedicated). Turbopuffer offers its Launch plan at $16 per month and Scale plan at $256 per month. Typesense Cloud's smallest cluster configuration costs $0.03 per hour, which Typesense says works out to $21.60 per month, and the first 720 hours are free.
Qdrant Cloud offers a free tier, and Milvus itself is listed with enterprise pricing available on request. For teams evaluating total cost of ownership, the self-hosted open-source options (pgvector, Qdrant, Weaviate, Vespa, LanceDB) eliminate software licensing fees but require investment in DevOps and infrastructure management.
The most cost-effective option depends heavily on your scale and operational capabilities. Small teams or startups may find managed services like Pinecone or Zilliz Cloud more economical when factoring in engineering time. Organizations with established infrastructure teams often find self-hosted options like Milvus, Qdrant, or pgvector deliver better value at scale.
When to Consider Switching
Several scenarios may prompt you to evaluate Milvus alternatives, each driven by different operational or technical needs.
If you want to eliminate infrastructure management, a fully managed service like Pinecone or Zilliz Cloud removes the burden of cluster provisioning, scaling, and maintenance. Zilliz Cloud is particularly appealing if you want managed Milvus compatibility without the operational overhead.
If you already run PostgreSQL, pgvector lets you add vector search capabilities without introducing a new database system. This consolidates your data stack, simplifies backups and monitoring, and lets you combine vector similarity search with traditional SQL joins and filters in a single query. pgvector works well for datasets up to tens of millions of vectors.
If you need multimodal data support, LanceDB is purpose-built for multimodal AI workloads, handling text, images, video, and point clouds in a unified system with native versioning and training pipeline integration.
If you need combined search capabilities, Typesense and Vespa both offer hybrid search that combines vector similarity with full-text keyword search. Weaviate similarly supports multiple search techniques within a single query, which can reduce architectural complexity when you need both semantic and keyword matching.
If operational simplicity is paramount, serverless options like Turbopuffer (built on object storage) or Pinecone (fully managed) minimize the infrastructure you need to maintain. These are strong choices when your team lacks dedicated database operations expertise.
If Rust-level performance matters, Qdrant is written in Rust and designed for high throughput with efficient memory usage, which can be advantageous for latency-sensitive workloads.
Migration Considerations
Migrating from Milvus to another vector database requires planning around data export, index rebuilding, and API changes.
Data export and format compatibility is the first hurdle. Milvus stores vectors in its own internal format, so you will need to export your vectors and metadata, then re-import them into the target system. Most vector databases accept vectors as arrays of floats, making the data transformation relatively straightforward. For pgvector, vectors are inserted using SQL syntax. For Pinecone or Qdrant, you would use their respective client SDKs to upsert vectors in batches.
Index rebuilding will be necessary regardless of the target database. Each system uses different indexing algorithms and parameters, and optimal settings vary by dataset. Plan for an initial period of index tuning and benchmarking on your actual data. pgvector recommends creating indexes after loading data for faster build times. Qdrant and Weaviate handle indexing automatically on data insertion.
API and SDK changes represent the most significant development effort. Milvus uses its own Python SDK (pymilvus) and gRPC API. Switching to pgvector means using SQL through any PostgreSQL client. Pinecone, Qdrant, and Weaviate each have their own REST APIs and language-specific SDKs. If you use LangChain or LlamaIndex, most of these databases have integrations that can reduce the migration effort at the application layer.
Moving to Zilliz Cloud is the lowest-friction migration path from self-hosted Milvus, since Zilliz Cloud is the managed version of Milvus built by the same team. Your existing pymilvus code and collection schemas should work with minimal changes.
Testing and validation should include recall benchmarking to verify that search quality is maintained after migration. Run your existing queries against the new system and compare result relevance before cutting over production traffic. Plan for a parallel-running period where both systems serve queries to validate consistency.