Milvus

Milvus started at Zilliz in Shanghai in 2019 and is now the default open-source answer when a team needs vector retrieval it can run itself. The licence is Apache 2.0 across the engine, with no seat limits and no field-of-use conditions, and the project is governed under LF AI & Data rather than by Zilliz alone. That distinction matters because Zilliz Cloud, the managed service, is a different product with a different engine underneath.

The 3.0 release, GA on 29 July 2026, is the second major rebuild in the project's history. Storage V3, codenamed Loon, moves to manifest-based columnar storage on object storage, and External Collection then lets you run search over Parquet, Lance, Iceberg or Vortex data in that object store without ingesting it into Milvus first. For teams whose embeddings are already produced by a lakehouse pipeline, that removes an entire copy step. The 3.0 line also rebuilt sparse indexing with SINDI and Block-Max WAND, so dense and BM25-style sparse retrieval can be combined in the engine rather than stitched together in application code.

Practicalities for anyone deploying from mainland China or building for Chinese users: the images are on Docker Hub and mirrored, the docs at milvus.io are fully English and current, and no mainland phone number or business licence is required for anything, since you are running your own binary. GPU indexing through CAGRA needs NVIDIA hardware and CUDA, with no Huawei Ascend path, so plan CPU HNSW or DiskANN if your cluster is domestic silicon. Milvus Lite covers prototyping in-process, Standalone covers a single node, and only the Distributed mode brings the etcd, message queue and multi-role complexity that people complain about.