Zilliz has launched Milvus 3.0, a significant overhaul of its open-source vector database that introduces lake-native data access and an enhanced retrieval engine for developers creating production AI applications. The update allows organizations to construct and deploy production-grade indexes over vector data that stays in object storage and open data formats, according to the company's announcement. The release expands the database's capabilities beyond conventional nearest-neighbor search to encompass richer ranking, aggregation, sparse search, and multi-vector retrieval by moving more query and retrieval operations into the engine itself.
The new version introduces External Collections, which define Milvus collections over data stored in Lance, Iceberg, Parquet, or Vortex formats. The system constructs vector, full-text, JSON, and scalar indexes over this data and makes them available through the same APIs as native collections, all without duplicating the source data into Milvus. Incremental synchronization maintains index currency as the underlying dataset evolves. Milvus 3.0 also debuts Loon, a manifest-based storage engine built to substantially cut the read amplification tied to low-latency point access on object storage, using Vortex—an open, Arrow-compatible columnar format—as its default. The release includes snapshots that generate point-in-time, read-only views of live collections at minimal incremental storage cost, letting offline jobs like evaluation, deduplication, and backfill validation run against a stable dataset while production writes proceed. A new Spark connector treats Milvus as a Spark DataSource V2, enabling Spark, Databricks, and EMR pipelines to read from and write to Milvus within standard batch workflows, while vector-native batch operators handle tasks including deduplication and clustering. The platform now supports server-side sorting and aggregation alongside faceted search, allowing applications to blend semantic relevance with attributes such as freshness, rating, price, category, tenant, and content source. For entities represented by multiple vectors—documents, images, or products—StructList delivers native multi-vector retrieval, letting a document retain vectors for multiple chunks or an image preserve vectors for multiple visual patches under a single entity and metadata record, with support for late-interaction retrieval models like ColBERT and ColPali. An optimized sparse index is approximately three times smaller while achieving comparable recall in internal testing, with additional enhancements including 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.
"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." The company stated that production AI systems typically keep one copy of data for real-time retrieval and another for offline processing, with data movement between the two creating additional storage overhead, lengthy exports, synchronization pipelines, and operational complexity. According to the vendor, the new capabilities provide a broader retrieval foundation for applications spanning RAG, semantic search, recommendation, multimodal discovery, agent memory, and AI data preparation.
The lake-native architecture addresses the core problem by permitting production indexing and retrieval to operate directly over data in object storage, eliminating the need to continually export, duplicate, and reconcile large datasets across separate systems. By bringing online retrieval and offline data improvement closer to a shared data foundation, the release reduces both storage duplication and pipeline complexity while enabling operations that previously required over-fetching data and processing it in application code to execute directly within the Milvus engine. The same technical foundation powers both open-source Milvus—which developers can deploy independently on Kubernetes or Docker, including in air-gapped environments—and Zilliz Cloud, a fully managed Vector Lakebase that extends the Milvus core with elasticity, operational simplicity, and enterprise capabilities for managing AI data at scale.
Milvus 3.0 is available under the Apache 2.0 license and remains a graduated LF AI & Data project, with support for Amazon S3-compatible object storage, Google Cloud Storage, and Azure Blob Storage. The release ships initially through the Python, Go, and Node.js SDKs, with Java support planned shortly after launch, while organizations preferring a managed platform can access Zilliz Cloud, which incorporates Milvus 3.0 as its core indexing and retrieval engine. The shift to lake-native architecture positions the database to handle increasingly complex AI workloads without forcing organizations to maintain parallel infrastructure for serving and analysis. For enterprises already invested in object storage ecosystems, the architecture may reduce the friction between experimentation and production deployment, though the real test will be whether teams can resist the operational inertia of duplicating data when millisecond-level latency becomes non-negotiable.