Embed 5: Frontier Embedding Models for Enterprise | Cohere

This post introduces Cohere's Embed 5 model family, featuring 'Pro' and 'Fast' tiers. Embed 5 Pro offers maximum retrieval quality across multimodal, multilingual, financial, code, and parsed-document data, achieving state-of-the-art scores on benchmarks like ViDoRe V3 and FinanceBench. Embed 5 Fast provides competitive performance for latency- and cost-sensitive workloads. Both tiers share a unified embedding space, enabling indexing with Pro and querying with either model. Key technical advancements include a 128K token context window, support for multimodal inputs (text, images, fused text-image), 100+ languages, and various output dimensions (2048 down to 256). Matryoshka representations and int8/binary embedding formats are introduced to reduce vector storage and search costs. The post also introduces RCP-nDCG@10 as a new retrieval methodology for performance evaluation. Performance gains are detailed across visually rich documents, financial datasets, parsed PDFs, fused text-image retrieval, page-image retrieval, and multilingual capabilities, with specific benchmark scores and comparisons to competitors like Voyage 4 Large, Gemini Embedding 2, and OpenAI text-embedding-3-large.