Cohere launches Embed 5: Pro for indexing, Fast for querying, no need to rebuild the vector database.

动察 Beating AI News: Cohere, an enterprise AI company invested in by Nvidia, has released a new generation of embedding models, Embed 5, available in two tiers: Pro and Fast. Embedding models are responsible for converting text and images into vectors for search, RAG, and Agent information retrieval. Both models support text, image, and mixed input, cover more than 100 languages, and have a maximum context length of 128K tokens.

The two models share the same vector space. Enterprises can first use the higher-quality Pro to index documents, then use the faster Fast to process queries, without needing to regenerate the entire set of vectors. Cohere's tests on 40 datasets show that the combined score of indexing with Pro and querying with Fast is equivalent to 98.4% of the all-Pro approach, a gap of about 1.6%.

In Cohere's tests, Embed 5 Pro averaged 85.8 on ViDoRe V3, while Fast scored 84.5, higher than Voyage 4 Large's 83.7 and Gemini Embedding 2's 83.2. Fast is priced at $0.08 per million text tokens, and Pro at $0.12.