Qdrant Releases a 10-Billion-Vector Benchmark for Large-Scale Search

Qdrant has released Qdrant-FineWeb-10B, a free, open-source vector-search dataset containing 10.07 billion vectors, alongside its new benchmarking framework, Supernova. The release was announced September 1, 2026, in the Hugging Face Blog, with infrastructure support from Vultr.

The benchmark is built from Hugging Face’s FineWeb corpus and uses the gte-multilingual-base embedding model. Its files include 24.47 TB of vectors plus 28.66 TB of source text and metadata. Qdrant says the dataset includes dense and sparse representations, filtered-search ground truth, and exact top-1,000 nearest-neighbor results for 100,000 queries.

Why it matters

Most vector-search benchmarks stop at roughly 10 million to 100 million embeddings, omit verified answers, or fail to represent hybrid systems that combine dense, sparse, multi-vector, and filtered retrieval. FineWeb-10B is intended to test systems under conditions closer to production deployments—and to expose performance differences that small, curated datasets can hide.

Generating the benchmark required more than one quadrillion distance calculations, performed in parallel on GPU hardware. That process also led Qdrant to create Supernova, an open-source engine for producing exact nearest-neighbor ground truth and evaluating databases at billion-vector scale without relying on proprietary tooling.

The release includes two additional datasets: PubMed-Multi-Vector, with 23.9 million dense and sparse vectors plus 8.37 billion multi-vector tokens, and Coyo-Vector-Embeddings, with 15.4 million 2,048-dimensional vectors covering text and images. For AI builders, the package offers a much tougher, reproducible way to compare retrieval architectures, indexing strategies, and hardware choices before deploying them in real applications.

Source: Hugging Face Blog