Zilliz, the company behind Milvus—an open-source vector database—recently announced the release of Milvus 3.0, a major architectural update that brings lake-native data access and a more expressive retrieval engine to developers building production AI applications.
According to the company, Milvus 3.0 enables organizations to build and serve production-grade indexes over vector data that remains in object storage and open data formats. It also moves more query and retrieval operations into the engine, expanding Milvus beyond traditional nearest-neighbor search to support richer ranking, aggregation, sparse search, and multi-vector retrieval.
This release also advances the shared technical foundation behind both open-source Milvus and Zilliz Cloud. Developers can deploy Milvus independently, while Zilliz Cloud builds on the Milvus core to deliver a fully managed Vector Lakebase spanning real-time serving, interactive discovery, batch analytics, and AI data operations.
“AI data is becoming larger, richer, and more dynamic, yet the systems used to serve and improve that data have remained fragmented,” said James Luan, co-founder and CTO of Zilliz. “Milvus 3.0 brings production retrieval closer to where data already lives and gives developers a more expressive engine for modern AI applications. That same foundation powers Zilliz Vector Lakebase, where we extend it with the elasticity, operational simplicity, and enterprise capabilities required to manage AI data at scale.”
Production AI systems commonly maintain one copy of data for real-time retrieval and another for offline processing. Moving data between the two introduces additional storage overhead, lengthy exports, synchronization pipelines, and operational complexity, the company said.
Milvus 3.0 introduces a lake-native architecture that allows production indexing and retrieval to operate directly over data in object storage.
Other features and capabilities include:
- External Collections define Milvus collections over data stored in Lance, Iceberg, Parquet, or Vortex. Milvus builds vector, full-text, JSON, and scalar indexes over the data and exposes them through the same APIs as native collections—without copying the source data into Milvus. Incremental synchronization keeps indexes current as the underlying dataset changes.
- Loon, the new manifest-based storage engine in Milvus 3.0, is designed to substantially reduce the read amplification associated with low-latency point access on object storage. It uses Vortex, an open and Arrow-compatible columnar format, as its default storage format.
- Snapshots create point-in-time, read-only views of live collections at minimal incremental storage cost. Offline jobs such as evaluation, deduplication, and backfill validation can run against a stable dataset while production writes continue.
- A new Spark connector exposes Milvus as a Spark DataSource V2, allowing Spark, Databricks, and EMR pipelines to read from and write to Milvus as part of standard batch workflows. Milvus 3.0 also introduces vector-native batch operators for tasks including deduplication and clustering.
Together, these capabilities bring online retrieval and offline data improvement closer to a shared data foundation, reducing the need to continually export, duplicate, and reconcile large datasets across separate systems, said Zilliz.
Additionally, Milvus 3.0 also expands retrieval beyond traditional top-K vector search.
New server-side sorting and aggregation, together with faceted search, allow applications to combine semantic relevance with attributes such as freshness, rating, price, category, tenant, and content source. Operations that previously required over-fetching data and processing it in application code can now execute directly within the Milvus engine.
For documents, images, and other entities represented by more than one vector, StructList introduces native multi-vector retrieval. A document can retain vectors for multiple chunks, an image can retain vectors for multiple visual patches, and a product can retain several multimodal representations under a single entity and metadata record. StructList also supports late-interaction retrieval models such as ColBERT and ColPali.
Milvus 3.0 further strengthens sparse and hybrid retrieval with an optimized sparse index that is approximately 3 times smaller while achieving comparable recall in internal testing. Additional enhancements include SINDI for learned-sparse vectors, server-side MinHash generation, nullable vector fields, custom dictionaries for full-text search, and broader support for Faiss-compatible indexes.
These capabilities provide a broader retrieval foundation for applications spanning RAG, semantic search, recommendation, multimodal discovery, agent memory, and AI data preparation, said the vendor.
Milvus 3.0 is available under the Apache 2.0 license and remains a graduated LF AI & Data project. It can be deployed on Kubernetes or Docker, including in air-gapped environments, and supports Amazon S3-compatible object storage, Google Cloud Storage, and Azure Blob Storage.
The release is available initially through the Python, Go, and Node.js SDKs, with Java support planned shortly after release.
For organizations that prefer a managed platform, Zilliz Cloud incorporates Milvus 3.0 as its core indexing and retrieval engine and extends it into a fully managed Vector Lakebase.
For more information about this news, visit https://zilliz.com/.