Zilliz Cloud

Zilliz was founded in Shanghai in 2017 and now runs out of Redwood Shores, California, which puts it in an awkward bucket: Chinese origin, American parent, Linux Foundation governance. Milvus, the engine underneath, was open-sourced in 2019 and donated to LF AI & Data a year later. It crossed 40,000 GitHub stars in late 2025 and is the usual answer when a team asks what handles a billion embeddings without falling over.

The managed product gives you Milvus without babysitting etcd, a message queue, object storage and four coordinator roles. Free tier covers 5 GB and five collections. Serverless bills by vector compute unit at roughly $4 per million, dedicated clusters start near $99 a month and storage pricing was cut steeply in January 2026. On the engine side, 2.6.x added JSON path indexing that Zilliz benchmarks at up to 100x faster metadata filtering, and 3.0 in July 2026 made the whole thing lake-native so indexes can serve directly over open formats in object storage.

Use it if retrieval is load-bearing for your product and you are past ten million vectors, or if you need dense, sparse and full-text ranking fused in one query instead of three systems stitched together. Skip it if you have two hundred thousand chunks and a Postgres box already running, because pgvector will do that job and you will not add a service to the on-call rotation. The self-host path matters most for manufacturing and public-sector work in China where the embeddings are not allowed to leave the building.