Source: qdrant.tech/blog/cohere-embed-5-qdrant
Cohere has released Embed 5, a family with a quality-focused Pro model (embed-v5.0-pro) and a lighter Fast model (embed-v5.0-fast). Both accept text and images, support more than 100 languages and share one embedding space. Qdrant wanted to see whether a single dense vector per PDF page could compete with late-interaction visual models such as ColPali and ColQwen, which usually store many vectors per page.
On five public English ViDoRe V3 corpora, covering 12,969 page images and 1,489 questions, Embed 5 Pro scored above several late-interaction models on mean nDCG@10. Qdrant notes that Cohere's training data is not published, so some of the results may reflect test data seen during training.
For storage, Qdrant combines Matryoshka truncation to 512 dimensions, 1-bit page vectors and 8-bit query vectors in Qdrant 1.19. It rescores the top 20 candidates with full 2048-dimension vectors kept on disk, which keeps 99.2% of full-vector quality at 64 bytes per page in RAM. In a separate test, reranking OCR text with Rerank 5 Pro raised mean nDCG@10 from 66.8 to 73.2.
Why it matters if you build agents
If you build PDF retrieval for agents, a single Embed 5 vector per page lets you search with ordinary HNSW indexes and skip OCR, though the recipe reads full vectors from disk during rescoring. Qdrant estimates about 610 MB of RAM for search vectors across 10 million pages, compared with about 76 GB for full float32 Embed 5 vectors.
These figures come from Qdrant's own tests on five English corpora, so check recall on your documents before relying on them in production.
Get stories like this every weekday
One short email with the day's agent tooling news.