# Milvus 2.6 made storage format and cost controls part of billion-scale retrieval claims Milvus announced version 2.6 on June 12, 2025, describing Storage Format V2, advanced search and architecture work intended for large vector collections. Canonical: https://rangoon.ai/insights/archive-milvus-2-6-billion-scale-retrieval-june-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-06-12 Tags: [Milvus](https://rangoon.ai/insights/tags/milvus/), [vector search](https://rangoon.ai/insights/tags/vector-search/), [retrieval](https://rangoon.ai/insights/tags/retrieval/), [AI infrastructure](https://rangoon.ai/insights/tags/ai-infrastructure/) Source note: Historical reporting: Milvus announced Milvus 2.6 on June 12, 2025 and described cost reduction, advanced search and architectural changes for large-scale vector workloads. Release notes later record the production v2.6.0 release and state that direct upgrades from pre-2.6.0 require the upgrade guide. Vendor performance claims depend on collections, indexes, filters, hardware and test method; they are not a universal billion-vector result. Image: Mixed vector and scalar collection passes through versioned storage and search layers. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Milvus 2.6 treated storage format as a retrieval-cost and operations decision. - The release combined search features with an architectural upgrade boundary. - Scale claims need corpus, indexing, filtering and hardware context. ## Milvus announced version 2.6 on June 12, 2025 [source 1](https://milvus.io/blog/introduce-milvus-2-6-built-for-scale-designed-to-reduce-costs.md) Milvus published its 2.6 announcement on June 12, 2025, presenting the release around lower cost, advanced search and changes for large-scale vector workloads. The article describes a direction toward collections with very large numbers of vectors, not an assertion that every installation contains a billion vectors or should use the same configuration. The dated post establishes the public announcement and the capabilities Milvus chose to foreground. Performance and cost language in a supplier release remains conditional on workload design. A system’s dimensions, update rate, index parameters, metadata filters, storage media and service objectives determine whether a feature produces a useful result in that system. ## Storage Format V2 made mixed scalar and vector storage an explicit architecture change [source 1](https://milvus.io/blog/introduce-milvus-2-6-built-for-scale-designed-to-reduce-costs.md) [source 2](https://milvus.io/docs/release_notes.md) Milvus 2.6 release material names Storage Format V2 among the new capabilities. Storage layout matters in a vector system because retrieval often combines vector similarity with scalar filters, point lookups and metadata access. A format change can alter compaction, reads, writes and resource use even when an application’s query text does not change. The production release notes say direct upgrades from earlier versions are not supported and direct users to an upgrade guide. That is an operational boundary worth keeping with the release: a major storage design is not a drop-in claim. It requires planning, compatibility review and a test against the actual collection lifecycle. ## The launch joined advanced search language to a stated cost objective [source 1](https://milvus.io/blog/introduce-milvus-2-6-built-for-scale-designed-to-reduce-costs.md) [source 2](https://milvus.io/docs/release_notes.md) The launch framed 2.6 around affordable vector search at large scale and described advanced search capabilities in the same release. Those phrases should travel with their measurement conditions. Search quality can change when a collection uses hybrid queries, different rerankers, scalar constraints, updated embeddings or a different recall target. Cost can change with index build time, compaction, storage tier, query volume and operations work, not only bytes per vector. The supplier’s framing identifies what the project intended to improve. It does not provide an independent benchmark for every model, corpus or deployment. A team comparing versions should preserve both relevance metrics and resource measurements rather than promote either one into a blanket conclusion. ## The v2.6 record makes upgrade state part of any before-and-after comparison [source 1](https://milvus.io/blog/introduce-milvus-2-6-built-for-scale-designed-to-reduce-costs.md) [source 2](https://milvus.io/docs/release_notes.md) Milvus documentation records v2.6.0 as a production release and calls out stability fixes and capabilities built on its release-candidate work. That separates three facts often blurred in software news: an announcement, a release candidate and a production version. It also means that later point releases should not be retroactively treated as features available on June 12 unless their own dates support that claim. For a retrieval team, the evidence needed for an upgrade decision includes the precise version, migration path, collection schema, index configuration and representative queries. The public launch and release notes explain the project’s declared scope; they do not verify a local migration or guarantee that an older client remains compatible. ## Sources 1. [Milvus: Introducing Milvus 2.6](https://milvus.io/blog/introduce-milvus-2-6-built-for-scale-designed-to-reduce-costs.md) 2. [Milvus 2.6 release notes](https://milvus.io/docs/release_notes.md) ## Related reading - [Qdrant 1.13 joined GPU indexing with collection-level operational limits](https://rangoon.ai/insights/archive-qdrant-1-13-gpu-indexing-strict-mode-january-2025/) - [Qdrant Cloud Inference put embedding generation, storage and indexing behind one API call](https://rangoon.ai/insights/archive-qdrant-cloud-inference-embedding-indexing-july-2025/) - [Pinecone made Assistant generally available with chat, context and document-format additions](https://rangoon.ai/insights/archive-pinecone-assistant-general-availability-january-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)