What is Milvus 3.0?
Milvus 3.0 is the newest major release of Milvus, the open source vector database hosted at LF AI & Data. It is the largest architectural update in the project’s history, introducing two core capabilities: lake-native retrieval, which lets Milvus index and search vector data that stays in object storage and open table formats, and a more powerful retrieval engine, which moves sorting, aggregation, faceted search, and multi-vector scoring directly into Milvus instead of application code.
Milvus 3.0 is available today under the Apache 2.0 license, maintained as an LF AI & Data project.
Why does Milvus 3.0 matter?
Most AI teams already store embeddings in a data lake: Parquet files, Lance tables, or Iceberg tables on S3, GCS, or Azure Blob Storage. Until now, searching that data meant either copying it into a vector database (creating a second copy and an ETL pipeline to keep in sync) or querying the lake directly (which forces a brute-force scan with no production-grade latency).
Milvus 3.0 removes that tradeoff. It lets teams build and serve vector indexes directly on data that never leaves the lake, while also absorbing more of the retrieval pipeline (sorting, counting, grouping, multi-vector search) into the engine itself.
What’s new in Milvus 3.0?
1. Lake-native infrastructure
- External Collections: Define a Milvus collection over data sitting in Parquet, Lance, Iceberg, or Vortex, and query it with the same APIs as a native collection. The source files never move, and Milvus builds vector, BM25, JSON, and scalar indexes over the external data.
- Loon (Storage v3): A manifest-based columnar storage engine purpose-built for the narrow point reads that follow an ANN search. In internal benchmarks, this cut measured I/O per point read from roughly 9.4 MB (Parquet baseline) to roughly 0.07 MB, about 135x less.
- Snapshots: Cheap, point-in-time, read-only views of a collection for evaluation, deduplication, or backfill testing, without copying the full dataset.
- Spark DataSource V2 connector: Lets Spark, Databricks, and EMR jobs read from and write to Milvus as part of standard batch pipelines.
- Online schema evolution: Add, fill, or drop collection fields while the collection keeps serving traffic, with no full rebuild required.
2. A more powerful retrieval engine
- Server-side ORDER BY: Sort filtered or searched results by scalar fields like price, rating, or timestamp inside Milvus, cutting client-side over-fetching.
- Aggregation and faceted search: Run count, sum, average, min, and max directly in Milvus, plus grouped facet counts after an ANN search.
- StructArray: Store a variable-length array of vectors (chunks, frames, tokens, patches) in a single row with one entity ID, with native support for late-interaction models like ColBERT and ColPali.
- SINDI and BM25 index compression: A new sparse retrieval algorithm plus a roughly 3x smaller sparse index at comparable recall, with up to roughly 10x the QPS on learned sparse embeddings such as SPLADE.
Who benefits most from Milvus 3.0?
This release is aimed squarely at teams building:
- AI agents whose underlying data changes constantly
- Multimodal or late-interaction retrieval (ColBERT, ColPali)
- Governed environments where data must remain in place
- RAG and knowledge-base systems over long documents
- Search that blends relevance with business signals like price, rating, or inventory
What comes next for Milvus?
The roadmap builds on the 3.0 architecture with predicate pushdown for External Collections, external backfill support, additional Spark operators, and support for more table formats, including Delta Lake and Apache Paimon.
Where can I get the full technical details?
The complete release write-up, including architecture diagrams, code samples, benchmark methodology, and the full feature table, is available on the Milvus blog: Announcing Milvus 3.0: Lake-Native Vector Search and a More Powerful Retrieval Engine.
How do I get started with Milvus 3.0?
- Read the release notes and quickstart, and get the source at github.com/milvus-io/milvus
- Join the Milvus Discord community or book a Milvus Office Hours session with the maintainers
About Milvus and LF AI & Data
Milvus is an open source vector database and a graduated project of LF AI & Data, the Linux Foundation umbrella for open source AI and data projects. Milvus 3.0 reflects continued investment from the Milvus maintainer community, including contributors who filed issues, tested release candidates, and shaped the feature set through real production feedback.
