October 2026 Qdrant 1.13 surge: myths vs facts for teams

Qdrant 1.16 improves filtered search via the ACORN algorithm to address quality drops when multiple low-selectivity filters exist. This algorithm examines neighbors of neighbors during HNSW traversal to prevent disconnected graphs when high cardinality filters reduce the search space. Qdrant 1.16 also provides inline storage, which places quantized vector data directly inside HNSW nodes. This feature allows the system to read both neighbor IDs and quantized vectors from consecutive disk pages during a single search iteration. This modification improves search speeds for low RAM environments. This works because quantization reduces the size of each vector, and without quantization, the original vector size exceeds the 4KB page capacity. Qdrant 1.16 also introduces tiered multitenancy to allow the combination of small and large tenants within a single collection through user-defined sharding, which prevents the noisy neighbor problem in shared instances. Qdrant uses its Rust implementation to deliver p99 latency under 50ms for many workloads. For users managing 1M vectors at 768 dimensions with 99% recall, Qdrant achieves approximately 1,800 queries per second on hardware with 8 vCPUs and 64GB RAM. This throughput exceeds self-hosted Weaviate on similar hardware by 40%.

Weaviate provides native hybrid search by merging BM25 keyword scores with dense vector similarity in one query. This process allows the database to blend semantic intent with exact word matches, which solves retrieval failures when users type exact product names or technical error codes. Weaviate’s module system can vectorize text during ingestion using models from providers like OpenAI or Hugging Face. Weaviate 1.37 released in April 2026 includes a native MCP Server so LLM agents can query the database directly. Weaviate’s GraphQL-based interface handles filtering, sorting, and aggregation at the database level to reduce retrieval plumbing. Weaviate requires a strict schema-first approach where every property needs a type before ingestion. This requirement makes metadata evolution difficult because every change necessitates a full schema migration. Pinecone provides a managed alternative that eliminates infrastructure management. Pinecone’s serverless tier costs $0.33 per million vectors stored and $8.25 per million queries. While Pinecone scales automatically, its metadata filtering imposes a 40KB limit per vector. If your retrieval depends on massive JSON payloads, this limit forces you to move metadata to an external store. You should examine how your specific query patterns handle these constraints.

Deployment costs for 10M vectors at 1536 dimensions vary by provider. A self-hosted Qdrant setup on a $40/month VPS with binary quantization handles these workloads. Weaviate Cloud for 10M vectors costs approximately $135/month when using Binary Quantization. Pinecone costs approximately $251/month for 10M vectors with 1M queries per day.

Teams using PostgreSQL can use pgvector for workloads under 5M vectors. pgvector uses HNSW indexing to provide high recall for these smaller datasets. At 5M vectors, pgvector p95 latency climbs to 80-140ms. For datasets exceeding 100M vectors, Milvus provides a distributed architecture with separate compute and storage layers. Milvus 2.6 replaced Kafka dependencies with a custom WAL system called Woodpecker. Milvus uses index types like IVF, HNSW, and DiskANN to manage large datasets. Milvus architecture includes specialized query, data, index, and proxy nodes. How will your team manage the migration if your vector requirements suddenly exceed the capacity of your existing relational database?