Qdrant
Rust vector search engine with native hybrid dense and sparse retrieval, self-hostable under Apache 2.0 or run as a managed cloud cluster.
Overview
Written in Rust and shipped as a single binary, Qdrant starts in Docker in about a minute or runs as a managed cluster on AWS, GCP or Azure. The distinguishing design choice is composability. Dense vectors, sparse vectors from BM25 or SPLADE++, payload filters and multi-stage reranking all compose inside one query API instead of being stitched together in application code, which is why teams doing retrieval over awkward enterprise documents tend to land here.
The 2026 releases have been aimed squarely at cost per vector. Turbo4 stores at 4 bits with no full-precision companion, roughly ninefold less storage, and memory tiers let you pin the HNSW graph in RAM while leaving payloads cold on disk. That combination is what makes billion-vector collections affordable on hardware you can actually rent. The tradeoff is stated plainly in the release notes: with no originals to rescore against, recall drops, so it is a knob for storage-bound workloads rather than a free win.
Model the economics before committing. Because Qdrant Cloud bills provisioned resources rather than reads and writes, a low-traffic index costs more here than on Pinecone serverless, and the comparison only turns in Qdrant's favour somewhere north of tens of millions of vectors. Self-hosting removes licence cost entirely under Apache 2.0 and replaces it with a rotation. Both are defensible. Picking by accident is not.
Key Features
- ✓ Version 1.19, shipped 5 August 2026, adds the Turbo4 datatype: 4-bit compression as the sole vector representation, roughly a ninefold storage reduction
- ✓ Memory tiers in 1.19 replace scattered cache settings with one parameter offering pinned, cached and cold modes across vectors, HNSW links, quantized data and payloads
- ✓ Native hybrid queries mixing dense vectors with BM25, SPLADE++ or miniCOIL sparse vectors in a single API call
- ✓ Tiered multitenancy from 1.16, which mixes small and large tenants in one collection and promotes heavy tenants onto dedicated shards
- ✓ ACORN filtered search, aimed at queries carrying several weakly selective filters at once
Where it holds
- • Fastest of the mainstream three in published 2026 benchmarks: around 6ms p50 on 1M vectors, against roughly 8ms for Pinecone and 12ms for Weaviate.
- • The self-hosted binary is the same engine as the cloud, so a laptop Docker container is a fair rehearsal for production behaviour rather than a toy.
- • Hybrid retrieval is built in rather than bolted on, and on messy enterprise corpora that matters more than the ANN benchmark does.
- • Raised $50M in March 2026 led by Advance Venture Partners with Bosch Ventures joining, past 250 million downloads. Abandonment risk is low.
Where it breaks
- • Cloud billing is resource-based, so you pay for provisioned RAM whether anyone queries or not, and the pricing page lists tiers without listing a per-hour rate.
- • At modest scale it loses on cost. Third-party 2026 comparisons put a 10M-vector cluster near $456/month against roughly $370 for Pinecone serverless, with the crossover only arriving around 50M.
- • Self-hosting past a single node means owning sharding, replication and rebalancing, and the new Turbo4 datatype buys its storage win by discarding the full-precision copy used for rescoring, which costs recall.
My Take
The 1.19 release on 5 August 2026 is the one worth reading. Turbo4 makes 4 bits the only vector representation instead of a compressed sidecar sitting beside a full-precision original, cutting storage around ninefold, and the new memory tiers collapse a pile of fiddly cache flags into pinned, cached or cold. Benchmarks published this year put it near 6ms p50 on a million vectors where Pinecone sits at 8ms and Weaviate at 12ms, though those margins compress once filters get aggressive. The bill is where it stings: resource-based cloud pricing charges for provisioned RAM, not queries, and the pricing page still declines to publish a rate.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Apache 2.0 core is free to self-host with no vector or node caps. Qdrant Cloud has a free forever tier at 0.5 vCPU, 1GB RAM and 4GB disk, a usage-based standard tier billed hourly on vCPU, RAM, disk and backup storage with a 99.5% SLA, and a premium tier carrying a minimum spend that adds SSO, private VPC links and a 99.9% SLA. Hybrid cloud and private cloud are quote-only. The pricing page publishes no per-hour rate; independent 2026 comparisons put a 10M-vector managed cluster near $456/month.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Qdrant holds and where it breaks. The fastest check is your own workload.
Visit website ↗Qdrant compared head to head
Alternatives to Qdrant
Milvus
open-sourceApache 2.0 vector database from Zilliz, self-hostable from a pip install up to a billion-vector cluster. The open engine under Zilliz Cloud, minus the proprietary Cardinal search core.
pgvector
open-sourcePostgres extension that adds vector types and approximate nearest neighbour indexes, so embeddings live in the same database as the rest of your data.
Zilliz Cloud
freemiumManaged vector database built on Milvus, the open-source engine behind most billion-scale retrieval systems, for teams who need hybrid search without running the cluster themselves.
seekdb
open-sourceAnt Group's OceanBase team open-sourced this AI-native search database under Apache 2.0 (seekdb, from the OceanBase 蚂蚁 stack). Vector, full-text and scalar filters in one SQL query, embeddable in...