Qdrant 1.13 surge and the competition for vector storage

Qdrant moves past the scale debate

Qdrant’s $28 million Series A funding and its binary quantization technology change the math for teams choosing between it and Weaviate or Milvus. This quantization reduces memory consumption by 32 times and increases retrieval speeds by 40 times because it compresses vectors into simple representations of zeros and ones. I see developers favor this approach for latency-sensitive workloads that require combining vector similarity search with structured attribute filters. The company provides managed cloud services on AWS, GCP, and Azure, alongside a managed on-premise edition for enterprises. You probably already know that managing infrastructure creates friction, but Qdrant’s hybrid cloud option lets you attach your own infrastructure to their managed service. The payload indexing system allows for complex filter conditions, such as finding vectors where category equals legal and date is after 2025-01-01, without scanning the entire index. Qdrant added hybrid search support in 2024, which helps close the gap with Weaviate’s capabilities.

Weaviate and Milvus performance trade-offs

Weaviate differentiates itself by combining dense vector search with BM25 sparse retrieval in a single query. This native hybrid search helps when queries contain both semantic intent and specific keywords like product codes or names. Weaviate uses a modular architecture to plug in different vectorization models directly at the database level. However, Weaviate requires more RAM than Qdrant or pgvector and has a steeper learning curve because of its schema design and GraphQL interface. Milvus targets billion-scale deployments through a distributed architecture that separates storage and compute. While Milvus handles massive datasets through distributed architecture, it requires Kubernetes and several dependent services, which makes the operational burden much higher than using a managed service like Pinecone or Qdrant Cloud. Milvus supports diverse indexing like IVF and HNSW, but managing it requires more hands-on work than Pinecone’s managed nodes. Zilliz, which commercializes the Milvus open source database, secured $60 million to support these massive scale requirements.

Scaling decisions and hardware requirements

If I use pgvector, I use my existing PostgreSQL infrastructure. pgvector uses the <-> operator for cosine distance and provides HNSW indexing to achieve recall rates above 95% for datasets up to 10 million vectors. For teams with a budget of zero, pgvector on a well-provisioned instance is the most cost-effective choice. However, pgvector performance degrades once a dataset reaches 100 million vectors. If I need to scale a RAG pipeline to billions of vectors, Milvus is the strongest candidate because its distributed design handles datasets that exceed the memory of a single machine. I would skip Weaviate if my priority is maximum raw speed for real-time recommendations, as Qdrant’s Rust implementation minimizes overhead. Does the memory savings from binary quantization compromise the accuracy enough for precision-heavy medical datasets?

Marqo targets developers who want to bypass third-party vector-generation tools from OpenAI or Hugging Face by providing an end-to-end system for vector generation, storage, and retrieval. Pinecone remains a strong option for teams that want zero operational overhead, though its per-query pricing becomes expensive for high-volume workloads. I recommend starting with pgvector if your scale is below 10 million vectors to avoid adding new services. For high-throughput, latency-sensitive applications, I choose Qdrant.