Qdrant
Listed at https://qdrant.tech
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Free 永久层:1 GB 存储 + 0.5M 向量,无限期;Cloud Standard 月付 $25 起(预付 $250/yr)按存储 $0.06/GB-月;Cloud Pro 月付 $80 起 按存储 $0.04/GB-月 + 高 IOPS;Dedicated 合同制 $2,500/月起 | per million tokens |
| Output | 按存储维度(GB)与可选性能包(Power Tiers)计费;无 API 调用费、无 per-vector 嵌入费;FastEmbed 按 token 数本地推理;BYO Embedding 由第三方 API 计费 | per million tokens |
🤖 Supported Models (9)
✨ Pros
- ✓Rust-based open-source vector database (Apache 2.0 + commercial QPL), production-grade performance with low memory footprint
- ✓Native hybrid search (Sparse + Dense Vectors) + Named Vectors (multiple models per collection)
- ✓FastEmbed with 9 built-in embedding models, local inference with zero extra cost
- ✓API Recommendations: auto-suggests vector index configs and HNSW parameters
- ✓GPU-accelerated indexing (100× speedup, 2026 GA); MMR / Discovery search for diversity
- ✓Cloud SOC 2 Type II + ISO 27001 + GDPR compliant; HIPAA (BAA) at Pro+ tier
⚠️ Cons
- ×High latency from mainland China (self-host OSS or BYOC for low latency)
- ×Cloud pricing lower than Weaviate Flex but $25 tier requires $250/yr prepaid
- ×FastEmbed model set slightly smaller than Weaviate Cloud's 8 but lighter weight
🎯 Best For
RAG apps needing Rust performance and low memory; GPU-accelerated large-scale (>10M vectors) indexing; hybrid search + named vectors for multimodal; BYOC for cloud cost control
💰 Pricing & Plans
| Plan / Tier | Price | Key Limits | Best For |
|---|---|---|---|
| Free Tier | $0 /mo, no credit card | 0.5 vCPU / 1 GB RAM / 4 GB disk; ~1M 768-dim vectors; unlimited API requests; community Discord support | Learning Qdrant, hobby projects, permanent proof-of-concept |
| Standard Tier | Usage-based, billed hourly | Dedicated resources (vCPU, RAM, storage billed per GB); 99.5% uptime SLA; 10x5 business-hours support; free tokens for paid inference models | Production workloads needing dedicated cluster with no long-term commitment |
| Premium Tier | Usage-based, billed hourly | Multi-AZ within a single region (3+ nodes); 99.9% uptime SLA (99.95% Multi-AZ); 24x7 critical incident response; SSO / SAML; customer-managed encryption keys; VPC private links | Regulated, high-availability workloads needing HIPAA and data residency |
| Hybrid Cloud | Contact sales | Data plane runs in your own Kubernetes cluster (AWS / GCP / Azure / on-prem); Qdrant operates control plane; you pay for underlying infrastructure | Data residency or compliance requiring vectors to stay in your infrastructure |
| Private Cloud | Contact sales | Dedicated isolated deployment; full control; on-premise or any cloud; custom SLA | Strict security or compliance needs requiring complete isolation |
| Cloud Inference (managed) | Per-token, model-dependent | Free models (BM25, all-MiniLM-L6-v2) carry no token charges; metered models billed per token with 5M free tokens/mo per model; external providers (OpenAI / Cohere / Jina AI / OpenRouter) proxied with your API key | Simplifying embedding pipeline — one vendor for vector DB + inference |
| BYOC (self-hosted OSS) | $0 (Apache 2.0 + QPL) | Full feature set; no vendor lock-in; you manage infrastructure; FastEmbed runs locally with zero API cost | Teams with existing ops capability wanting full control and zero licensing cost |
🔧 API & Developer Experience
- •Official clients in six languages: qdrant-client for Python (pip install qdrant-client[fastembed]), @qdrant/js-client-rest for TypeScript/JavaScript, qdrant-client for Rust, go-client for Go, Qdrant.Client for .NET, and io.qdrant/client for Java — all maintained by Qdrant and version-locked to server releases.
- •REST API as primary interface: All interaction takes place via the REST API (OpenAPI spec at api.qdrant.tech). gRPC is available on port 6334 for binary-protocol performance, with one-to-one method parity for every REST endpoint.
- •FastEmbed: built-in local inference: pip install qdrant-client[fastembed] gives you 9 embedding models (BGE small/base/large, MiniLM-L6-v2, E5 base/large, Jina v2 small multilingual, CLIP ViT-B-32) running entirely on your machine — zero API keys, zero per-token cost, ONNX Runtime under the hood.
- •Cloud Inference: in-cluster embedding: Managed Cloud clusters ship with Cloud Inference enabled by default. Replace the pre-computed vector in your upsert or query with an Inference Object (text or image + model name) and Qdrant embeds, stores, and searches in one call — no external API hop.
- •One call for hybrid + multimodal search: Native hybrid search fuses dense, sparse (BM25 / SPLADE++), and multi-vector (ColBERT) representations in a single query. Pass an image Inference Object and the same call returns ranked multimodal results.
- •MCP server for agentic tools: The Qdrant MCP Server (github.com/qdrant/mcp-server-qdrant) exposes collection search, point retrieval, and upsert as MCP tools, so Cursor, Claude Desktop, and other MCP clients can query your vectors directly.
- •Payload filtering with HNSW integration: Filters on payload attributes extend the HNSW graph itself, so filtering happens during the graph traversal — a single-pass operation rather than pre- or post-filtering. ACORN search further improves accuracy when combining multiple high-cardinality filters.
🔍 Native Hybrid Search, BM25 & GPU-Accelerated Indexing
Qdrant is a Rust-built, open-source vector database that treats hybrid search as a first-class primitive rather than a bolt-on feature. Version 1.20 introduced BM25 full-text search that runs natively alongside dense vectors — no sparse vector model needed, no separate Elasticsearch cluster to synchronize. You fuse lexical and semantic scores in one query, and the same engine handles multi-vector representations like ColBERT for late-interaction retrieval. For bulk loads, GPU-accelerated HNSW indexing delivers up to 4x faster index construction on AWS, while scalar quantization cuts RAM by 4x with only 1-2% recall loss and TurboQuant pushes compression to 8-32x. The result is a single Rust binary that can serve semantic, keyword, and multimodal search from one collection with production-grade latency.
🌐 China Access & Latency
Qdrant Cloud has no mainland-China region. Managed Cloud runs on AWS and GCP in regions like Frankfurt, North Virginia, and Sydney, so a China-facing workload must route cross-border and accept 200-400ms latency on the data path. Measured from an Alibaba Cloud Malaysia host, qdrant.tech and cloud.qdrant.io are CloudFront-fronted and answer in roughly 30-80ms, while the actual gRPC data path to a managed cluster in Frankfurt or North Virginia takes considerably longer. For production inside mainland China, self-hosting is the practical answer: the OSS engine is Apache 2.0 licensed (QPL for the commercial-friendly bits), runs as a single Rust binary or Docker container, and deploys cleanly on Alibaba Cloud, Tencent Cloud, or on-premise Kubernetes for near-domestic latency. FastEmbed and client-side inference run locally with zero external API calls, so a self-hosted Qdrant deployment in China needs no cross-border network access at all.