Qdrant

license

Apache-2.0(OSI, engine core)

source

public(Rust)

managed-tier

Qdrant Cloud (hosted, paid + free tier)

core-gated

ungated(no ee/ path no second license and zero repo-wide hits for enterprise or license key. The two engine features the pricing table marks absent from OSS are both in the open tree - GPU indexing as the lib/gpu crate and resharding across four lib/collection modules with eight consensus tests - so what is sold is control plane rather than engine)

Nothing in the engine is held back, so this scores 5. The managed Qdrant Cloud tier sitting on top is not a gate under this rubric, and a read of the repository and the pricing page finds no gating mechanism anywhere: a code search across the whole repository for `enterprise` returns zero hits and one for "license key" returns zero, there is a single Apache-2.0 LICENSE at the root, and there is no ee/, enterprise/ or second-licensed path in the tree. The pricing table is where a careless read goes wrong. It marks GPU Indexing and Shard Splitting as absent from the open source build, and both are in the open engine: GPU indexing is its own crate at lib/gpu (Cargo.toml, src/, nvidia_icd.json) beside segment, quantization and sparse, and resharding is implemented and consensus-tested at lib/collection/src/collection/resharding.rs, lib/collection/src/shards/resharding.rs, lib/collection/src/shards/shard_holder/resharding.rs and lib/collection/src/shards/transfer/resharding_stream_records.rs, with eight tests/consensus_tests/test_resharding*.py files. What the table actually compares is the managed offering: uptime SLAs, managed backup and disaster recovery against the "custom automation" you would run yourself, multi-AZ topology, disk encryption with customer keys, and Enterprise SSO on the Qdrant Cloud console. Hybrid Cloud is described as running "managed Qdrant clusters on your own infrastructure" - the same engine, operated by someone else. A managed service sold beside a complete open core is not a gate.

• https://github.com/qdrant/qdrant/blob/master/LICENSE recorded 2026-06-04

  Apache-2.0 license text

• https://github.com/qdrant/qdrant recorded 2026-06-04

  v1.18.2 (4 Jun 2026); Qdrant Cloud managed tier with free option

• https://api.github.com/repos/qdrant/qdrant/contents recorded 2026-08-12

  Root tree: one LICENSE beside Cargo.toml, config, docs, lib, openapi, pkg, src, tests and tools. No ee/, enterprise/, commercial/ or proprietary/ path and no second license file, so no part of the engine ships under different terms.

• https://api.github.com/repos/qdrant/qdrant/contents/lib recorded 2026-08-12

  The engine's crates: api, blobstore, bm25, collection, common, edge, gpu, macros, posting_list, quantization, segment, shard, sparse, storage, trififo, uio-grpc-client, wal. GPU indexing is the gpu crate, published here, which is the first of the two features the pricing table marks as absent from OSS.

• https://qdrant.tech/pricing/ recorded 2026-08-12

  Read for a mechanism and found none. Its OSS-versus-paid comparison is control plane: no uptime SLA for OSS against 99.5% free tier and 99.9% (99.95% multi-AZ) on Standard/Premium, "custom automation" backups against managed backup and disaster recovery, multi-AZ topology, disk encryption with custom keys on AWS/GCP, and "Enterprise SSO Authentication" on Premium+. It also marks GPU Indexing and Shard Splitting as absent from OSS - both are in the open tree, so those rows describe the managed offering rather than the engine. Hybrid Cloud is "Run managed Qdrant clusters on your own infrastructure"; Private Cloud is a "Dedicated, isolated deployment".

Banded on downloads of the Python client: the registry reports 19,077,155 downloads of qdrant-client in the last month, well clear of the >10M floor that level 5 takes on this map (for scale, pydantic-ai sits near 31M and langgraph near 58M at the same level). That is one level above Milvus on download volume, and Qdrant is a leading open vector database for retrieval in RAG and agent stacks. The figure is for the Python client rather than the server, which is what this product declares as its countable artifact - a server has no download channel of its own, so its client's downloads are the best available usage signal.

• https://pypistats.org/api/packages/qdrant-client/recent recorded 2026-08-13

  last_month = 19,077,155 downloads for qdrant-client (last_week 4,271,948, last_day 681,811).

Level with Milvus, which is the 4 rung on the storage feature matrix, which bands on how much of the persist-index-retrieve path a product provides itself, at what scale and over how many retrieval modes: a complete durable system, distributed by sharding and replication, covering dense, sparse and multi-vector retrieval with configurable score fusion. It does not reach the 5 rung, which is for a platform that hosts and evaluates ranking or embedding models inside the serving path; score fusion and late interaction over vectors the caller supplies are not that. The headline throughput comparison is vendor-run and published at qdrant.tech/benchmarks. ANN-Benchmarks corroborates inclusion and nothing more - it lists 38 algorithms, Qdrant and Milvus (Knowhere) among them, and publishes only interactive recall-versus-queries-per-second plots with no ranking table - hence medium confidence. Rebanded on this matrix when the product moved out of agent_tools_protocols; the band did not move.

• https://ann-benchmarks.com/ recorded 2026-08-18

  Qdrant listed among 38 benchmarked ANN algorithms, alongside Milvus(Knowhere), faiss-ivf, scann, hnswlib and weaviate. Results published as interactive recall-versus-queries-per-second plots per dataset; no ranking table.

• https://qdrant.tech/benchmarks/ recorded 2026-08-18

  Vendor benchmark, Observations section - "Qdrant achieves highest RPS and lowest latencies in almost all the scenarios, no matter the precision threshold and the metric we choose. It has also shown 4x RPS gains on one of the datasets." Compared against Milvus, Weaviate, Elasticsearch and Redis.