Milvus 3 cuts vector point-read I/O 135x

Zilliz released Milvus 3.0 with a vendor benchmark showing its new Loon storage path cut point-read I/O roughly 135x, from 9.4 MB with Parquet to 0.07 MB with Vortex across 3 million 128-dimensional vectors and 256 concurrent readers. The Milvus 3.0 launch details say read-only External Collections can index Parquet, Lance, Iceberg, or Vortex data in object storage without copying it into Milvus; the release also adds snapshots, Spark DataSource V2, server-side aggregation, and multi-vector retrieval. For builders doing the retrieval work described in the modern AI engineer’s stack, the architecture could remove a serving copy and its synchronization pipeline—but the 135x result is internal, so benchmark latency and cloud I/O costs on your own access pattern before migrating.