Qdrant vs Weaviate

Qdrant and Weaviate are the two vector databases in our catalog, and on a VPS they look alike: one container each, REST and gRPC APIs, a volume for the data, and almost nothing in use until you load vectors into them. The split is in what each expects you to bring. Qdrant stores and searches vectors you computed elsewhere, and gives you a dashboard, payload filters and three kinds of quantization to do it well. Weaviate can compute the vectors for you by calling a model provider at import time, and layers hybrid search and role-based access on top. The licences differ too, in a way that matters only if you want the enterprise parts.

Who computes the embeddings

Qdrant is a search engine for vectors you already have. Its Quickstart creates a collection with a vector size and a distance metric and then upserts points; every example passes the arrays in. FastEmbed, the embedding library Qdrant maintains, is "a lightweight Python library built for embedding generation" that runs on the ONNX runtime, which means it runs in your application, not in the Qdrant server. Hybrid search is there, but it is also vector-in: the Query API takes dense and sparse vectors as prefetch sub-requests and fuses them with RRF, which "considers the positions of results within each query", or DBSF, which "keeps the raw scores from each query but normalizes their distributions before combining".

Weaviate will do the embedding step for you. With a vectorizer module configured, "Weaviate automatically generates embeddings for objects that are added to the database", which it does "by providing the source data to the integration provider, which then returns the embeddings to Weaviate". The provider list is long: OpenAI, Cohere, Google, Mistral and Hugging Face among the API-based ones (Anthropic is a generative provider only), and Ollama and text2vec-transformers among the locally hosted; the Docker guide runs the transformers case as a second container, slower without a GPU. Generative and reranker modules sit beside the vectorizers for RAG. Hybrid search here is keyword plus vector from one input: it "combines the results of a vector search and a keyword (BM25F) search by fusing the two result sets", balanced by alpha, where "An alpha of 1 is a pure vector search".

If your pipeline already calls an embedding model, Qdrant fits it without asking anything else of you. If you would rather hand the database raw text and an API key, Weaviate removes a service from your stack.

Filtering, compression, and the dashboard

Both filter on structured data alongside the vector search. Qdrant's payload filters combine must, should and must_not clauses over match, range, geo and full-text conditions. Weaviate stores objects beside their vectors and filters on their properties.

Both compress vectors to save memory. Qdrant documents scalar quantization, "achieving 4x compression", binary quantization "for up to 32x compression", and product quantization for "up to 64x compression". Weaviate lists rotational quantization as recommended, plus product, binary and scalar.

Qdrant has a web UI. The dashboard at /dashboard is "an intuitive and efficient graphic interface for your Qdrant Collections, REST API and data points", with a console to "Run HTTP-based calls". Weaviate ships no web interface; you talk to it through a client library or curl.

Security and access

Neither is safe on a public port as shipped. Qdrant is blunt about it: "Self-hosted open source deployments are not secure by default and are not production-ready." You set QDRANT__SERVICE__API_KEY and clients send it in the api-key header; a read_only_api_key gives read-only access; and granular JWT-based keys with per-collection read or write permissions need jwt_rbac: true. "TLS must be used to prevent leaking the API key over an unencrypted connection." Weaviate says anonymous access is "strongly discouraged except for development or evaluation purposes" and offers API keys through AUTHENTICATION_APIKEY_ENABLED and its sibling variables, or any "OpenID Connect" issuer with discovery. On top sits RBAC: root has "full access to all resources", viewer "read-only access to all resources", and you define further roles as bundles of permissions. Our snippets set an API key on both and turn Weaviate's RBAC on with a root user.

Backups differ in shape. Qdrant takes per-collection snapshots. Weaviate backs up "at instance or collection granularity" through a module you enable: backup-s3, backup-gcs and backup-azure are the ones recommended for production, and backup-filesystem is single-node only.

Licences

Qdrant is Apache-2.0, the whole repository.

Weaviate's LICENSE says code "is licensed under the BSD-3-Clause License" except that "Within the "wl" directory and all its sub-directories, source code is Copyright Weaviate BV" and "is only available under a separate enterprise license". "Enterprise features are enabled via license key", and "The BSD license does not grant the right to use these features or circumvent any license key." Everything a self-hoster reaches in this comparison is in the BSD part; the wl directory is the line to know about before you rely on an enterprise feature.

Footprint and setup

Both are one container, one volume. Qdrant publishes REST on 6333 and gRPC on 6334 and keeps data under /qdrant/storage; Weaviate publishes REST on 8080 and gRPC on 50051 and keeps data at /var/lib/weaviate. Neither publishes a RAM minimum, because sizing depends on how many vectors you store and at what dimension, so our 1 GB floors are estimates for a small collection. In our install check (GCP e2-standard-2, Ubuntu 26.04, October 2026), Qdrant idled at ~17 MB of RAM and ~278 MB of disk; Weaviate idled at ~43 MB of RAM and ~292 MB of disk. Those are empty-database numbers; your vectors, not the engine, will decide the server. We rate both 2 out of 5 to deploy.

Which to choose

Choose Qdrant if you compute embeddings in your own code and want a small Apache-2.0 engine with a dashboard to look at the collections, payload filters you can index, and quantization you can dial from 4x to 64x.

Choose Weaviate if you want the database to call the embedding model for you, keyword and vector search fused in one query, and roles for more than one team on the same instance, and you are content with BSD-3-Clause for the core and a licence key for the wl extras.

Running Qdrant or Weaviate on a VPS

Both installs are verified on Ubuntu 26.04; the pinned compose files with API keys set are on the Qdrant and Weaviate pages. Both speak plain HTTP and h2c, so put a reverse proxy with TLS in front of every port you expose, and firewall the gRPC port to the clients that need it. If what you are really sizing is an LLM app that uses one of these underneath, Dify ships Weaviate in its default compose and carries a Qdrant profile; see Dify vs Langflow for that layer.