Cohere launches embed 5, outperforming competitors in key benchmarks

Cohere has launched Embed 5, a new embedding model family designed for enterprise search, RAG, and agentic retrieval. The model is available in two tiers: Pro and Fast. Pro prioritizes retrieval quality, while Fast focuses on latency and cost efficiency. Both models support text, images, and fused text-image inputs, covering 100+ languages and handling up to 128K tokens. They share the same embedding space, allowing indexing with Pro and querying with Fast. Embed 5 is currently available via Cohere’s API, Model Vault, Microsoft Foundry, and Amazon SageMaker, with private VPC and on-prem options through vLLM. Embed 5 outperforms competitors like Voyage 4 Large, Gemini Embedding 2, and OpenAI’s text-embedding-3-large in benchmark tests. On ViDoRe V3, Embed 5 Pro scores 85.8, while Fast scores 84.5. In finance benchmarks, Pro leads with scores of 80.1, 90.0, and 85.0, with Fast ranking second. Multilingual results show mixed performance, with Gemini Embedding 2 surpassing Pro in several languages. Cohere’s new RCP-nDCG@10 metric, which evaluates reranking quality, is used for these comparisons, though independent validation remains pending. The model employs Matryoshka representation learning and lower-precision outputs to reduce storage and improve efficiency. Cohere recommends 1024-dim int8 as the default setting, balancing quality and cost. Embed 5’s ability to process images and fuse them with metadata enhances its utility for scanned documents and visual data.