Dino10B: first 10-billion-scale nearest-neighbor search benchmark, 10.3TB of DINOv3 embeddings
philipvollet · x · 2026-11-12
Researchers including FAISS authors Swann Bessa, Matthijs Douze, and Hervé Jégou released Dino10B, the first clustering/nearest-neighbor benchmark at 10-billion-vector scale. It contains 10B 1024-dim uint8 DINOv3 ViT-L/16 patch embeddings from Creative Commons images (10.3TB), organized as 16 nested subsets from 100k to 10B vectors, each a prefix of the next. It ships with 100k held-out queries and exact top-10 ground truth at every subset size, plus 99M separately held-out training queries — enabling reproducible, scaling-aware evaluation of ANN search and clustering algorithms.
More from Infra
- AMD runs MolmoAct2 robot policy locally on a Ryzen AI Max+ 395 mini PC — DJiafei · 2026-11-12
- Opinion: AI infra is becoming a coding runtime as intelligence units outgrow tokens — prateekj · 2026-11-12
- PrefixRelay: open-source proxy gives Cursor 97%+ KV cache hits on local models — jjusko20 · 2026-11-12
- Kinetic, a Rust inference engine, answers questions on a 22.6k-token doc in 0.13s vs llama.cpp's 10.95s — jasonkneen · 2026-11-12
- HP starts selling $132,000 GB300 desktop with 748GB memory for local AI — gnukeith · 2026-11-12
- Dev runs out of GitHub Actions minutes, shows off Mac Studio with unreleased M5 Ultra chip — ow · 2026-11-12