Evals · conare-embed-1

ViDoRe v3, embedding

Search over 19,252 business document pages with 2,419 questions in English and French.

conare-embed-1

67.0

nDCG@10, higher is better · #1 of 9

  1. 1conare-embed-167.0
  2. 2Perplexity pplx-embed-v2-late-9b, page imagesOur run of the open weights66.5
  3. 3Perplexity pplx-embed-v2-late-9b, page textOur run of the open weights66.1
  4. 4Mixedbread search, page images64.2
  5. 5Mixedbread search, page text62.3
  6. 6voyage-4-largeMixedbread's run60.6
  7. 7gemini-embedding-2Mixedbread's run58.7
  8. 8voyage-multimodal-3.5, page imagesMixedbread's run57.5
  9. 9BM25Mixedbread's run49.7

Accuracy against list price

56606468$0.05$0.10$0.20$0.50$1.00$2.00up and to the left is betternDCG@10price per 1M tokens, log scaleconare-embed-1 · 67.0 · $0.10voyage-4-large · 60.6 · $0.12gemini-embedding-2 · 58.7 · $0.20Mixedbread, page images · 64.2 · $1.50Mixedbread, page text · 62.3 · $1.50
List prices per 1M tokens. Mixedbread publishes no embedding price; its $1.50 is indexing per 1M content tokens. Perplexity's open 9B has no API price, so it isn't plotted.

nDCG@10, higher is better · read 14 October 2026

ViDoRe v3 was one of the benchmarks used to choose conare-embed-1. Perplexity's 9B on page images is level with it within the 95% interval.

How it was measured

ViDoRe v3, its 8 public datasets and 2,419 questions in their original English or French, searched over the page text the benchmark publishes; nDCG@10 on the human labels, each dataset weighted equally. Mixedbread's rows are its published per-question results; the gaps carry a 95% interval from a paired bootstrap over questions (10,000 draws).

See it on your own data.

Book a demo