Zilliz has released Milvus 3.0, introducing a lake-native architecture that enables production AI retrieval directly from open data formats while expanding the capabilities of the Apache 2.0-licensed open-source vector database.
Zilliz has released Milvus 3.0, a major architectural upgrade to the world’s most widely adopted open-source vector database, introducing a lake-native architecture that enables production indexing and retrieval directly from object storage without copying data. Released under the Apache 2.0 licence, Milvus continues as a graduated LF AI & Data project.
The release allows developers to build indexes over data stored in open formats including Lance, Iceberg, Parquet and Vortex through External Collections. Milvus 3.0 also introduces Loon, a new manifest-based storage engine that uses the open, Arrow-compatible Vortex format to reduce read amplification for low-latency access to object storage.
Beyond traditional nearest-neighbour search, Milvus 3.0 adds server-side sorting, aggregation, faceted search, sparse and hybrid retrieval, and native multi-vector retrieval through StructList, supporting late-interaction models such as ColBERT and ColPali. Additional enhancements include an optimised sparse index, SINDI learned sparse vectors, server-side MinHash generation, nullable vector fields and broader Faiss-compatible index support.
The release also introduces Snapshots for point-in-time datasets and a Spark DataSource V2 connector for Spark, Databricks and EMR workflows. Milvus 3.0 supports deployment on Kubernetes or Docker, including air-gapped environments, with SDKs for Python, Go and Node.js available initially.
“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.”
