Qdrant vs Weaviate vs Milvus vs Pinecone: Vector Databases for RAG (2026)

Every RAG stack we have built on these pages depends on a vector database. When we self-hosted Dify, its Compose stack quietly pulled in Weaviate. When we ran Langflow and Open WebUI with Ollama, the first serious question after "it works" was "where do the embeddings live?" In 2026, four names dominate that shortlist: Qdrant, Weaviate, Milvus, and Pinecone.

This is a mixed field, and that shapes the comparison. Three are open source (Apache-2.0 or BSD-3-Clause) and can be self-hosted with a quick docker run. Pinecone is different: a closed-source managed service with no repository or self-hosting path. That is the product, not a criticism — but star counts, quickstarts, and licence comparisons only apply to three of the four, and we flag it wherever it matters.

A ground rule before we start: every available performance benchmark is vendor-run, so we won't be crowning a speed winner. More on that below.

Developer traction: the three you can actually star

Milvus leads on raw stars, with LF AI & Data foundation governance and the Zilliz ecosystem behind it. Qdrant's pitch is single-binary simplicity; Weaviate's smaller count belies the quickest release cadence in the field. Treat all three as traction signals, nothing more.

What each one actually is

Qdrant — the Rust engine you run with one command

Qdrant (Apache-2.0) is a vector search engine written mostly in Rust (87% of the repository, per the GitHub API). Its own docs describe HNSW-indexed search over dense or sparse vectors, payload (metadata) filtering with indexed fields, and distributed scaling via sharding and replication, over HTTP and gRPC with six official clients. Its hybrid Query API fuses dense and sparse results with RRF or DBSF scoring, and its docs detail four quantization methods reaching up to 64x compression. The Python client even runs a local in-process mode (QdrantClient(":memory:")) for prototyping with no server at all. Latest release: v1.18.3, 17 Jul 2026, with a minor line roughly every two to three months. The company announced a $50M Series B led by AVP on 12 Mar 2026 — a useful stability signal, by its own account.

Weaviate — the AI-native modular database

Weaviate (BSD-3-Clause, 96% Go) calls itself an "open-source, AI vector database". Its main feature is hybrid search, fusing vector similarity with BM25F keyword scoring through a single alpha knob and two fusion algorithms. Its multi-tenancy is distinctive — one shard per tenant, with states ACTIVE, INACTIVE, or OFFLOADED to S3, and a documented claim of a million concurrently active tenants on around 20 nodes. Deployment spans Weaviate Cloud, Docker, Kubernetes, and an evaluation-grade "Embedded Weaviate" mode. Cadence is the fastest here: v1.38.6 shipped on 21 Jul 2026, and three minor lines all received patches in July alone. Its $50M Series B, led by Index Ventures, dates to April 2023.

Milvus — the scale-out heavyweight with foundation governance

Milvus (Apache-2.0) is the only one of the four under neutral foundation governance: created by Zilliz, donated to the LF AI & Data Foundation, graduated June 2021. Its docs pitch "a cloud-native architecture designed for billion-scale or even larger scenarios", with search, insertion, and indexing scaling independently over separated storage and compute. It is also the only one with documented GPU acceleration, including NVIDIA's CAGRA index. Deployment spans Milvus Lite (a Python library — pip install pymilvus and you have a vector database in your notebook), Standalone (single-machine Docker), and Distributed (Kubernetes). Multi-tenancy comes in four documented strategies, up to "millions" of tenants via partition keys. Latest stable is v2.6.20 (14 Jul 2026); a 3.0 beta is tagged as of July 2026. Zilliz, its commercial arm, reported $113M in total funding as of its 2022 Series B extension.

Pinecone — the closed SaaS with no docker run

Pinecone is proprietary, closed-source, and managed-only. There is no public repository and no self-host path; the product is the API. Its serverless architecture separates write and read paths, stores vectors in distributed object storage, and partitions data into namespaces. It ships integrated inference — hosted embedding and reranking models — plus full-text search in public preview since May 2026. There are no version numbers, just a changelog with multiple updates a month. Corporate note: in September 2025 Ash Ashutosh was appointed CEO, with founder Edo Liberty moving to Chief Scientist; the last disclosed round is a $100M Series B from 2023.

Side by side

Self-hosting the open three: the official quickstarts

These are the vendors' own Docker quickstarts, quoted from their documentation as of 21 Jul 2026 — reproduced for reference, not an executed tutorial. All three run comfortably next to the Dify, Langflow, and Open WebUI stacks from our earlier guides.

Qdrant: one image, one command, data persisted locally, dashboard included.

docker pull qdrant/qdrant
docker run -p 6333:6333 -p 6334:6334 \
    -v "$(pwd)/qdrant_storage:/qdrant/storage:z" \
    qdrant/qdrant

Port 6333 serves REST (with the dashboard at localhost:6333/dashboard) and 6334 serves gRPC. Per the docs, all data lands in ./qdrant_storage under the default configuration.

Weaviate is also a one-liner, pinned to a version tag:

docker run -p 8080:8080 -p 50051:50051 cr.weaviate.io/semitechnologies/weaviate:1.38.3

Port 8080 is HTTP, 50051 is gRPC. One flag deserves attention before real data arrives: the default ships with AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED=true, and Weaviate's own docs say anonymous access "is strongly discouraged except for development or evaluation purposes." Harden it before it leaves your laptop. Docker Compose is supported for fuller configurations — which is how Dify bundles it.

Milvus standalone is heavier: the official path is a lifecycle script that provisions the server plus an embedded etcd and volumes.

curl -sfL https://raw.githubusercontent.com/milvus-io/milvus/master/scripts/standalone_embed.sh -o standalone_embed.sh
bash standalone_embed.sh start

The server listens on 19530, a WebUI sits at http://127.0.0.1:9091/webui/, and etcd takes 2379; the same script handles restart, upgrade, stop, and delete. If that is too much for an experiment, Milvus Lite gives you the same API from pip install pymilvus — positioned for notebooks and edge devices, and per the docs "only suitable for small scale vector search use cases."

With Pinecone, there's no quickstart to paste. It's closed-source SaaS; you sign up and call the API.

Managed pricing: four very different menus

All four vendors offer a free tier, but the terms are different enough to require a close look. Figures are from each vendor's published pricing as of 21 Jul 2026.

Pinecone prices in tiers: its pricing page lists a free Starter tier (up to 2 GB storage, 2M write and 1M read units a month), and its release notes add Builder — a flat $20/month plan for individual developers introduced 4 May 2026. Above those sit Standard at a $50/month minimum on pay-as-you-go usage (storage $0.33/GB/mo; writes $4–$4.50 and reads $16–$18 per million units, varying by region and cloud), and Enterprise at a $500/month minimum with a 99.95% SLA. Minimums are monthly, not annual.

Weaviate Cloud offers an always-free sandbox (100k objects, 1 GB memory, 10 GB disk, one collection, up to three tenants), then Flex serverless from $45/month on monthly pay-as-you-go, and Premium from $400/month on a prepaid contract. List prices vary by cloud provider and region.

Zilliz Cloud (managed Milvus) leads with a generous free cluster: 5 GB of storage — enough for a million 768-dimension vectors, by its own docs — 2.5M vCUs a month, up to five collections, no card required, plus $100 in credits on work-email signup (credits expire after 30 days). Beyond that it sells on-demand, dedicated, and pay-per-operation serverless billing. Exact list rates are not published in a form we could verify, so we won't quote them.

Qdrant Cloud has a free-forever single-node cluster (0.5 vCPU, 1 GB RAM, 4 GB disk) with limited region choice, then usage-based pricing billed hourly. There is no published list entry price — the pricing page is a calculator, and Hybrid and Private Cloud are sales-quoted. If procurement needs a headline dollar figure, Qdrant currently makes you ask.

Data control and the ASEAN angle

Only two of the four publish a region list that includes Southeast Asia. Zilliz Cloud lists AWS ap-southeast-1 and Google Cloud asia-southeast1, both Singapore — no Azure SEA region, no Jakarta, and region support "can vary by workload type" per its docs. Pinecone added AWS ap-southeast-1 (Singapore) for serverless indexes on 1 May 2026, on Builder, Standard, and Enterprise — the only SEA/APAC region on its list, with the docs warning that a serverless index's cloud and region cannot be changed after creation.

Qdrant Cloud and Weaviate Cloud do not publish region lists for their standard tiers — Qdrant's docs say only "a broad choice of regions" across AWS, Azure, and Google Cloud, and Weaviate publishes regions only for its Premium plan ("Any region on AWS, GCP, or Azure", by request). We can't tell you whether a Singapore region exists on those tiers, because the vendors haven't said. For the open-source three, the clean answer is self-hosting: an Apache-2.0 or BSD-3 engine in your own SG, MY, or ID datacentre — or on the same VPS as your Dify stack — is the one residency route no vendor roadmap can take away. Pinecone has no equivalent: residency there means choosing the Singapore region and trusting the service.

How to choose

Qdrant is the choice for the simplest self-host: one image, two ports, a bundled dashboard, and a clear quantization story for RAM. The local :memory: Python mode means prototype and production share one API. The trade-off: its managed cloud has no published list price to budget against, and no published SEA region.

Choose Weaviate if hybrid search ergonomics and SaaS-style multi-tenancy are your core problem: the single-alpha BM25F+vector knob is the cleanest hybrid interface here, and shard-per-tenant with S3 offload suits products with many small customers. It is also what Dify hands you by default — just fix that anonymous-access setting.

Choose Milvus if you are planning for scale you can name in billions, want GPU indexing, or prefer foundation-governed open source over single-vendor projects. Milvus Lite makes the entry ramp gentle; Distributed on Kubernetes is the documented path up. The heavier standalone script is the price of the architecture.

Choose Pinecone if you want zero operations and accept closed source as the cost. The $20 Builder tier is the cheapest published managed entry, a Singapore region is live, and integrated embedding and reranking remove two services from your stack. You trade away self-hosting, code inspection, and exit ease — vectors move out less freely than a Docker volume does.

FAQ

Which vector database is fastest for RAG?

Nobody can honestly tell you from public data: every available benchmark is vendor-run (see the honesty box above). Run your own corpus through the open-source harnesses on your own hardware — it is an afternoon's work and the only number that transfers to production.

Can I start locally with no server at all?

Yes, two good ways: Qdrant's Python client runs in-process with QdrantClient(":memory:") or a local path, and Milvus Lite arrives with pip install pymilvus — both share their server APIs, so code migrates unchanged. Weaviate's Embedded mode is evaluation-only; Pinecone has no local mode.

Which options keep data in Singapore?

With published guarantees: Zilliz Cloud (AWS and GCP Singapore regions) and Pinecone (AWS ap-southeast-1, since May 2026). Qdrant and Weaviate don't publish standard-tier region lists, so the verifiable route there is self-hosting in-region.

Is Milvus really open source if Zilliz runs the cloud?

The engine is Apache-2.0 and governed by the LF AI & Data Foundation (graduated June 2021) — the only one of the four under a neutral foundation. Zilliz created it and sells the managed version. Qdrant and Weaviate are open source too (Apache-2.0 and BSD-3-Clause), but their repos remain company-controlled.

Do all four work with LangChain and LlamaIndex?

Yes — all four publish official integration docs for both frameworks, so the orchestration layers in our Dify and Langflow guides can target any of them. Framework support won't be your deciding factor; self-hosting, pricing, and residency will.