LanceDB vs Milvus & Zilliz

LanceDB is an end-to-end multimodal lakehouse for curation, feature engineering, training and search on one table. Milvus is an open-source vector database, and Zilliz Cloud is its managed service.

Thank you Name Surname
Your submission has been received successfully.
We’ll get back to you as soon as possible.
In the meantime, please check your email — we’ve sent you a confirmation.

Short answer

LanceDB is the end-to-end data layer for multimodal AI, and Milvus is a vector database. Both are open source under Apache 2.0, so the choice comes down to scope. LanceDB is an open, multimodal AI lakehouse, while Milvus is a vector database that Zilliz also runs as a fully managed service, Zilliz Cloud.

LanceDB Enterprise connects curation, feature engineering, training access and retrieval through the same versioned tables, reducing duplicate datasets and synchronization between separate stores. LanceDB tables store raw images, video and audio next to their labels, embeddings and features, in the open Lance format that Spark, Ray, DuckDB and PyTorch can read.

Milvus is a capable vector database with many index types including DiskANN and GPU indexes, dense, sparse and binary vectors, and deployments from an embedded library to clusters of tens of billions of vectors. Zilliz Cloud adds a free cluster and published list prices.

At a glance

LanceDB vs Milvus & Zilliz: feature comparison

FeatureLanceDBMilvusZilliz Cloud
What it isMultimodal lakehouse for AIOpen-source vector databaseManaged Milvus service
ScopeCuration, feature engineering, training and search on one tableVector database, laptop to distributed systemsManaged Milvus for vector search applications
LicenseApache 2.0 (OSS); commercial (Enterprise)Apache 2.0Managed service; AUTOINDEX is proprietary
DeploymentEmbedded library in OSS; managed or BYOC on AWS, GCP and Azure in EnterpriseLite (embedded), Standalone (Docker), Distributed (Kubernetes)SaaS or BYOC on AWS, GCP, Azure
StorageS3, GCS, Azure Blob or local disk; NVMe cache (Enterprise)Object storage; query nodes load data into memory by defaultDedicated object storage buckets; tiered-storage clusters
Extra services to runNone for LanceDB OSS; runs inside your appetcd for metadata, MinIO for object storage and Woodpecker as the message queue, by defaultNone on SaaS; Zilliz operates them
Data formatOpen Lance format, read by Spark, Ray, Trino, DuckDB, Polars, PyTorch, and moreStorage V3 writes Parquet or Vortex (beta, opt-in)Import from Parquet (recommended), JSON or NumPy; read through Milvus SDKs and REST
Images, video, audioRaw bytes in binary columns; lazy-loading Blob API—Volumes hold data files for import and migration
Training data loadingPyTorch Dataset; StreamingDataset with shuffle and resume——
Feature engineeringEmbedding functions on insert; Python UDFs with distributed backfills (Enterprise)Function module calls embedding providers at insert (beta); BM25 and MinHash functions backfill existing rows (3.0)Embedding and reranking via model providers; managed backfill jobs (AWS, on request)
New embedding modelNew column; raw media columns aren't rewrittenAdd nullable vector field (2.6.18+); NULL until upsertedAdd nullable vector field; backfill job fills existing rows (AWS, on request)
Vector indexIVF family (PQ, RQ, SQ, FLAT, HNSW variants); managed automatically in EnterpriseHNSW, DiskANN, IVF, SCANN, GPU (CAGRA) and moreAUTOINDEX (proprietary); other Milvus indexes on request
Vector typesDense, binary, multivector; multiple vector columnsDense, sparse, binary; multiple vector fields; embedding lists (3.0)Dense, sparse, binary; embedding lists (StructArray, beta)
Full-text and hybridBM25; hybrid with RRF by defaultBuilt-in BM25 function; RRF or weighted rankerBuilt-in BM25; semantic, lexical and hybrid
FilteringSQL filters on vector, full-text, hybrid or plain queries; LIKE, regex, nested fields; scalar indexes; pre- or post-filterScalar filter expressions incl. LIKE and regexFilter expressions incl. LIKE and regex; JSON and array operators; Milvus scalar indexes
Query interfacePython, TypeScript, Rust SDKs; Java (Enterprise); SQL endpoint (Enterprise)SDKs (Python, Java, Go, Node.js, C++) and RESTMilvus SDKs and REST
VersioningEvery write is a version; time travel, tags, branchesCollection snapshots (3.0 beta)Backups (Dedicated clusters); snapshots (private preview)
TransactionsACID transactions——
PricingOpen source under Apache 2.0 for OSS; contact LanceDB for EnterpriseApache 2.0, self-hostedServerless from $0/mo; Enterprise from $197/mo
ComplianceSOC 2 Type II, HIPAA, GDPR (Enterprise)—SOC 2 Type II, ISO 27001, GDPR, HIPAA

What is the difference between LanceDB and Milvus?

LanceDB is a multimodal lakehouse, while Milvus is a vector database. LanceDB keeps curation, feature engineering, search and training attached to one versioned table instead of moving data through separate systems. Milvus describes itself as a vector database that runs from a laptop up to large distributed systems.

LanceDB OSS is an Apache 2.0 library that runs inside the application, and LanceDB Enterprise is a distributed platform. Milvus was created by Zilliz and is licensed under Apache 2.0. Zilliz Cloud is Zilliz's fully managed Milvus service, and Milvus's own docs point users who want managed Milvus there.

Terminology

Milvus stores entities in collections, and each entity has scalar and vector fields. The closest LanceDB terms are table (collection), row (entity) and column (field). A Milvus collection can have several vector fields, as a LanceDB table can have several vector columns. LanceDB also uses the word namespace, for the catalog path that groups tables.

How do LanceDB and Milvus architectures compare?

LanceDB Enterprise and Milvus both keep data on object storage and scale compute separately, while LanceDB OSS runs inside the application with no separate server.

LanceDB Enterprise query nodes, cache-backed plan executors and indexers scale independently, with table data and indexes in object storage and hot data in a distributed NVMe cache. By default, Milvus must load a collection into query-node memory before it can be searched, and the loaded data must stay under 90% of the query nodes' total memory. An optional Tiered Storage mode, in beta since 2.6.4, loads field data on demand instead.

Which is better for ML training data, LanceDB or Milvus?

LanceDB is built as the data layer between raw multimodal datasets and model training, while Milvus describes itself as a vector database and Zilliz Cloud as a managed Milvus service for vector search applications. A LanceDB table holds raw images, video and audio in binary columns next to labels, embeddings and features. Zilliz Cloud's managed volumes are a staging store for data files that get imported or migrated into a cluster.

A LanceDB table implements the PyTorch Dataset contract, so it can be passed straight to a PyTorch DataLoader. For larger jobs, LanceDB's StreamingDataset adds prefetching, deterministic shuffles, resumability and multithreaded transforms, and works with both LanceDB OSS and LanceDB Enterprise on local disk or object storage.

A training run can read only the rows that match a filter and only the columns it needs. Since every write creates a table version, data versions and tags can tie a model checkpoint back to the exact rows and features it was trained on.

Search is part of curation in this workflow. LanceDB uses filters, vector search and full-text search to find hard negatives, long-tail failure modes, duplicate clusters and low-quality samples.

How do feature engineering and embedding model changes compare?

LanceDB and Milvus can both add a vector field for a new embedding model to an existing table or collection, and both then need the new embeddings written for existing rows. In LanceDB, embeddings are columns, so a new model's embeddings go into a new column next to the existing ones, without rewriting the raw media. In Milvus 2.6.18 and later, an added vector field is NULL for existing entities until the application writes vectors through upsert or a backfill workflow.

Branches are available in both LanceDB OSS and LanceDB Enterprise. They let teams run experiments and backfills without disturbing production reads on main.

LanceDB Enterprise feature engineering

Feature engineering with LanceDB Functions, which run Python UDFs, is available only in LanceDB Enterprise. In both editions, new columns can be derived from existing data with SQL expressions, and the embedding function registry computes embeddings on insert.

  • A backfill runs a UDF as a distributed job to fill a column, and each batch is checkpointed so partial results survive job failures.
  • Columns marked for automatic backfill are recomputed by a managed agent when data changes.
  • Chunkers expand one row into many, for example splitting videos into clips or chunking documents.
  • Materialized views manage expensive UDF columns and can be used to optimize data layouts for training.

LanceDB schema changes can also alter and drop columns, and don't usually rewrite existing data, though data type changes can take more work.

Milvus

Milvus's Function module, in beta in 2.6.x, can call embedding providers such as OpenAI, AWS Bedrock and Google Vertex AI to embed text fields at insert time. Milvus's docs say to add a user-defined vector field when the application generates the embeddings and writes them to Milvus.

  • Added fields must be nullable, and vector search skips entities whose new vector is still NULL.
  • Milvus does not support changing a field's data type in any version.
  • Schema changes other than the supported add and drop operations require recreating or migrating the collection.

How do search features compare?

LanceDB, Milvus and Zilliz Cloud all support dense vector search, BM25 full-text search, hybrid search and reranking. LanceDB combines these capabilities with open, versioned tables that support production retrieval, data curation, feature engineering and model training.

  • LanceDB offers IVF and IVF_HNSW index families, with options including PQ, SQ and RaBitQ quantization, and can build its vector indexes on GPUs, automatically in LanceDB Enterprise. Milvus offers a wide range of vector index types, including HNSW, DiskANN, SCANN and GPU indexes such as GPU_CAGRA.
  • LanceDB applies SQL expressions across vector, full-text and hybrid queries, including nested fields, pattern matching and regular expressions. LanceDB supports scalar indexes and pre-filtering or post-filtering. Milvus uses its own filter expressions on scalar fields, which also support LIKE and regular expressions.
  • Both support Reciprocal Rank Fusion and weighted ranking for hybrid search, and reranking. LanceDB offers built-in reranker integrations for Cohere, Jina, Voyage AI and cross-encoders, plus custom rerankers. Since Milvus 3.0, rerankers can be composed into a Function Chain inside one search request.
  • LanceDB stores raw media, labels, embeddings and features together in versioned tables. Teams can retrieve production results, identify hard negatives and curate training examples from those tables. A Milvus collection stores entities with scalar and vector fields for search.

LanceDB makes retrieval part of a shared data foundation for both production applications and model development. Milvus focuses on vector search from a laptop up to large distributed systems.

How do LanceDB, Milvus and Zilliz Cloud scale?

Milvus positions Standalone for up to 100 million vectors and Distributed for up to tens of billions. Milvus states that it scaled to tens of billions of vectors in 2023. Zilliz's documented limit is in compute units, with up to 32 CUs on a Standard Dedicated cluster and up to 2,048 on Enterprise.

LanceDB OSS handles millions of vectors on a single node, and LanceDB Enterprise is the recommended edition for hundreds of millions. LanceDB Enterprise is built for workloads of 100B+ rows, and Enterprise customers regularly operate on billions of rows. In a published benchmark over 10 billion vectors on 10 nodes, LanceDB Enterprise reached 21 ms p99 for top-100 search using IVF with RaBitQ and 20 partitions examined per query. Distributed search in LanceDB Enterprise is listed as coming soon. In LanceDB Enterprise, query serving, read execution and index building scale independently, and in LanceDB's tests, automatic GPU indexing built indexes over billions of rows in under four hours on a 1–8 GPU cluster.

What do Milvus and Zilliz Cloud do well?

Milvus's docs cover a wide range of indexes and deployment sizes.

  • Milvus supports many vector index types, including DiskANN and GPU indexes such as GPU_CAGRA.
  • Milvus runs as an embedded Lite library, a single Docker image or a distributed Kubernetes cluster positioned for up to tens of billions of vectors.
  • Milvus supports multi-tenancy at the database, collection, partition and partition-key levels, with partition keys documented for millions of tenants.
  • Milvus has official Python, Java, Go, Node.js and C++ SDKs.

Can you self-host LanceDB or Milvus?

LanceDB OSS and Milvus can both be self-hosted. Milvus has three modes, Milvus Lite (an embedded Python library), Milvus Standalone (one Docker image) and Milvus Distributed (on Kubernetes). Milvus Lite supports Ubuntu and macOS and is recommended for up to a few million vectors. A Standalone instance can't be upgraded online to a Milvus cluster.

LanceDB OSS runs in-process inside a binary, container or serverless function. LanceDB Enterprise can be installed into the customer's own cloud account (BYOC) on AWS, GCP or Azure. Zilliz Cloud BYOC is generally available and runs the data plane in the customer's own VPC.

Since LanceDB OSS and LanceDB Enterprise use the same Lance format, moving between them needs no data conversion. Zilliz supports migrating from self-hosted Milvus to Zilliz Cloud by connecting to the instance or uploading backups.

How does compliance compare?

LanceDB Enterprise and Zilliz Cloud both list SOC 2 Type II, HIPAA and GDPR, and Zilliz also lists ISO/IEC 27001. On Zilliz's pricing page, HIPAA eligibility is a Business Critical plan feature. Zilliz Cloud encrypts stored data with AES-256, and customer-managed keys are a Business Critical feature. LanceDB Enterprise encrypts data at rest in both object storage and its cache. LanceDB can encrypt S3 data at rest with the customer's own AWS KMS key, and LanceDB Enterprise on Azure uses Azure Private Link.

Which should you choose?

Choose Milvus & Zilliz if

  • You only need vector search, and want DiskANN or GPU search indexes such as GPU_CAGRA.
  • Your retrieval relies on sparse vectors from models such as SPLADE.
  • You only need search, and want it as a managed service with a free cluster, a $0 Serverless start and published list prices.
  • You need ISO/IEC 27001 certification on a managed service.
  • Your services need an official Go or C++ SDK.

Choose LanceDB if

  • You train models on the same multimodal data you search, and want one table for raw media, labels, embeddings and features.
  • You want to feed training jobs directly from the table, with shuffling, filtering and resumable loading.
  • You are building production search or agents over open multimodal tables.
  • You add features or try new embedding models often, and want LanceDB Enterprise to backfill the new columns from data already in the table.
  • You need to tie each model checkpoint back to an exact data version.
  • You want the data in an open format you own, which Spark, Ray, DuckDB and PyTorch can read.

Frequently asked questions

Is Milvus open source?

Yes. Milvus is distributed under the Apache 2.0 license. Zilliz Cloud is Zilliz's managed Milvus service, and its AUTOINDEX index type is proprietary. LanceDB OSS is also Apache 2.0.

What is the difference between Milvus and Zilliz Cloud?

Milvus is the open-source vector database, and Zilliz Cloud is Zilliz's fully managed Milvus service. Zilliz Cloud runs as SaaS or BYOC and adds a proprietary AUTOINDEX.

Can I use LanceDB as a PyTorch data loader?

Yes. A LanceDB table works as a PyTorch Dataset, so it can be passed straight to a DataLoader. For larger jobs, LanceDB's StreamingDataset adds shuffling, prefetching and resumable loading, and works with both LanceDB OSS and LanceDB Enterprise.

What happens when I change embedding models in Milvus?

Since Milvus 2.6.18, a new vector field can be added to an existing collection, and it must be nullable. Existing entities have NULL in the new field until the application writes vectors through upsert or a backfill workflow, and vector search skips them until then. In LanceDB, a new model's embeddings go in a new column without rewriting the raw media.

Does Milvus need Kubernetes?

Only Milvus Distributed does. Milvus Lite is an embedded Python library and Milvus Standalone is a single Docker image. Milvus Distributed is installed on Kubernetes, for example with the Milvus Helm charts.

Can LanceDB replace Milvus?

For vector, BM25 full-text and hybrid search, LanceDB covers the same core features, and it also stores raw media and feeds training jobs from the same table.

How much does Zilliz Cloud cost compared to LanceDB?

Zilliz lists Standard from $0 per month on Serverless and Enterprise from $197 per month on Dedicated. LanceDB OSS is Apache 2.0.

Try LanceDB

Start with LanceDB OSS for search and training data on your laptop, or talk to us about LanceDB Enterprise for curation, feature engineering and training data at scale.