Cohere’s two new embedding models share the same vector space, process 128,000 tokens, support over 100 languages and images, and cost $0.12 and $0.08 per million tokens respectively.
On September 30, Cohere presented Embed 5, a family of embedding models for enterprise document search. It is available in two versions: Embed 5 Pro, which aims for the best search quality, and Embed 5 Fast, which is lighter, for high-volume, low-latency queries. Their particularity: a common vector space, which allows a corpus to be indexed with Pro and then queried with one or the other without recalculating the index.
Multimodal and multilingual
Both models accept up to 128,000 tokens, text, images or both combined, in more than 100 languages. Vectors can be produced in floating, int8 or binary, on 256 to 2,048 dimensions. On the ViDoRe V3 document review, Pro scores 85.8 average and Fast 84.5, ahead of Voyage 4 Large (83.7). Cohere also claims first place in several financial tests (FinanceBench, FinQA) and strong gains in Farsi, Telugu and Hindi compared to Embed 4.
Pricing: $0.12 per million text tokens for Pro, $0.08 for Fast, $0.40 for images. Models are available through the Cohere API, Microsoft Foundry, Amazon SageMaker, and in private deployment with vLLM.
Read the full article on the original source
Go to the source →