Cohere embed-v4

A high-performance embedding model optimized for multilingual retrieval and 96% storage cost reduction via advanced quantization.

Embed-v4 maps text into a 1024-dimensional vector space to power high-accuracy retrieval-augmented generation (RAG). It supports over 100 languages and processes 512-token sequences. This iteration focuses on compression-aware training: developers use int8 or binary quantization to slash storage costs by 96% while maintaining performance. It ranks at the top of MTEB leaderboards for complex datasets (legal documents and technical manuals). Native support for Pinecone and Weaviate ensures rapid deployment for production-ready AI applications.

https://cohere.com/embeddings