Zilliz, the company behind the open-source vector database Milvus, is unveiling Milvus 2.6, the latest iteration of its database technology designed to accommodate the modern demands of AI workloads without a modern price tag. Born from the feedback of users running large-scale AI applications, Milvus 2.6 reduces costs across storage, compute, operations, and developer time while introducing a variety of operational efficiencies.
With Milvus 2.6, Zilliz recognizes the increasing costs associated with deploying sophisticated AI, where vector search has become intrinsically tied to use cases such as semantic search, recommendations, and retrieval-augmented generation (RAG). At its core, this release is about lowering the barrier of entry to vector search, reducing infrastructure costs while forwarding a new, innovative, diskless architecture.
“As AI continues to advance, enterprises face mounting compute and storage demands. But few organizations have the luxury of unlimited budgets,” said Charles Xie, CEO and founder of Zilliz. “Milvus 2.6 is designed to let teams scale their vector workloads while keeping resource usage—and costs—under control. Innovations in this release help avoid the typical tradeoff between performance and affordability, making high-scale AI deployment accessible for more teams and industries.”
This release’s central theme—lowering infrastructure and storage bills—is achieved through the following capabilities:
- Tiered storage with hot/cold data separation, automatically moving frequently accessed vectors to high-performance storage while relocating less-used data to more economical options; works seamlessly with Cohesity, Pure Storage, MinIO, and NetApp storage providers
- Int8 vector compression for HNSW indexes which stores dense vectors with 8-bit integers, reducing memory requirements while maintaining search accuracy
- RabitQ 1-bit quantization, capable of pushing quantization to an extreme to achieve comparable retrieval quality with only half memory cost
- Write-ahead log (WAL) with Woodpecker which eliminates the need for external message queues such as Kafka or Pulsar, paired with a diskless architecture that reduces costs while enhancing write performance
“For organizations migrating from legacy search engines or general-purpose databases, Milvus 2.6 offers better performance at dramatically lower cost—up to 8x reductions reported in real-world benchmarks. Our goal is to continue to drive costs down in future releases,” said Xie.
Aside from 2.6’s obvious cost efficiencies, the database’s new diskless architecture helps Milvus eliminate a common pain point of complexity and expense—external message queues. As Xie explained, “This architectural shift improves write throughput, reduces latency, and simplifies deployment—especially for real-time or geo-distributed applications.”
The new diskless architecture comes with the following to further improve operational efficiency:
- Streaming node component for real-time data ingestion built directing into the Milvus platform, eliminating external message queues on the write path
- CDC (change data capture) with BulkInsert simplifies data replication across instances in different geolocations
- Storage v2 format which is optimized for performance and future compatibility with data processing frameworks such as Apache Spark
- APT/YUM deployment streamlines installation and upgrades, reducing operational overhead
Additionally, Milvus 2.6 helps boost developer productivity with more built-in tools to the platform. These capabilities include:
- Data-in, data-out approach which enables direct ingestion of raw content (such as text, images, and audio) with built-in inference capabilities, eliminating the need for external pre-processing pipelines
- Custom reranker that allows developers to apply custom scoring logic via scalar fields and user-defined functions
- Text and JSON search, driving support for advanced text processing capabilities such as advanced tokenization for Asian languages (Japanese/Korean); JSON path/flat/key indexing; and match and phrase queries
- Sampling and aggregation queries for faster iteration loops during development
“Milvus 2.6 isn’t just a performance upgrade—it’s a comprehensive rethink of how vector search can be deployed and scaled in the real world,” Xie noted. “These innovations make Milvus 2.6 one of the most cost-effective and scalable vector databases available today—open source, transparent, and trusted by a global developer community.”
To learn more about Milvus 2.6, please visit https://zilliz.com/.