ViDoRe V3 covers about 26,000 pages and 3,099 queries across ten domain corpora (eight public, two private), with human page-relevance, bounding boxes, and reference answers in six languages. It sits on the MTEB visual-document track. Vendor tables that quote a single average (for example Cohere Embed 5's 85.8 / 84.5) may use parsed text and a custom metric such as RCP-nDCG@10 rather than the public first-stage protocol — check the evaluation recipe before comparing models.