Weaviate 1.40 ships namespaces as generally available, if you have a license

Weaviate published v1.40.0 on 7 October 2026. The GitHub release is not a prerelease, and the breaking-changes section says none. Namespaces are labeled generally available. So is dropping a vector index. Reindexing a property is marked Preview. HFresh work, a search REST API, and 4-bit rotational quantization are in the notes without those labels.

Namespaces, and the license gate

Under “Namespaces (GA)” the notes say: “Namespaces add control-plane and data isolation between users on a shared cluster.”

They do not name a cloud plan or a price. They do list pull request 13298, “Make namespaces a feature requiring license,” merged on 30 September 2026. The pull request says the namespace handler is gated on a valid license. “Without a license all namespaced endpoints return a 403 error with an explanation.” The check runs in open-source code and calls into licensed code only when a valid license is provided. “Servers with namespaces enabled will still boot without a license and can still create backups etc but is not really usable.”

Multi-tenancy is a separate control. The multi-tenancy docs say it provides data isolation, that each tenant is stored on a separate shard, and that data in one tenant is not visible to another. The 1.40 notes describe namespaces as control-plane and data isolation between users on a shared cluster. They do not say namespaces replace multi-tenancy.

GA, Preview, and unlabeled sections

“Alter Schema - Drop vector index (GA)” is the next heading. The line under it says: “Support for dropping inverted indices from existing properties, allowing users to reclaim disk space by removing indices that are no longer needed.” The pull requests under that heading are about dropping a vector index, including a MUVERA bucket and an HFresh directory. The feature is marked GA. The sentence under the heading says inverted indices.

“Alter Schema - Reindex property (Preview)” says it “Adds support for changing property’s index types.” That is the section marked Preview.

The HFresh section lists performance work on memory, disk writes, and allocations. The heading is not labeled Preview or GA. “HFresh MUVERA” has one pull request, “Decouple Routing and Rerank Budgets,” also unlabeled. The vector index docs say multi-vector embeddings on HFresh were added in v1.40, and only with MUVERA. A query is encoded with MUVERA and searched like a single vector. Survivors are then scored with MaxSim against the original token vectors.

“Search REST API” says it “Adds dedicated REST API for performing search queries.” The pull requests add hybrid, near-object, aggregate, and near-vector endpoints, and one drops an experimental flag so the family can be on by default. The heading is not marked GA or Preview.

RQ4 is titled “4Bit Rotational Quantization” and lists centered quantization. It is not marked Preview. The RQ docs say 4-bit RQ was added in v1.40 and “stores each dimension in 4 bits, which gives a smaller index than 8-bit RQ at some cost in accuracy.” It requires HNSW. A dynamic index uses it only after converting to HNSW. Centering, which the docs date to v1.39.3, is valid on HNSW with bits set to 4 and is not supported for multi-vector embeddings without MUVERA. The 1.40 notes still list a centered-quantization pull request.

One pull request also gates deduplicated backups on a Weaviate license key. Weaviate’s release-notes page lists v1.40.0 as the latest release, still lists 1.39 and 1.38 as supported, and says 1.37 and older are no longer maintained.