Cohere Embed 5

Cohere's fifth-generation embedding models for enterprise search, RAG and agentic retrieval, in two tiers that share one embedding space: Embed 5 Pro for maximum quality and Embed 5 Fast, a lighter model for latency- and cost-sensitive query paths with on average 2.4× Pro's document throughput. Because the space is shared, a corpus indexed with Pro can be queried with Fast without re-indexing; across 40 development datasets the mixed pairings lose on average 1.6% (Fast queries) and 2.7% (Pro queries) against same-model baselines. Both tiers take text, images or fused text-plus-image inputs, cover 100+ languages, accept up to 128K tokens, and output Matryoshka embeddings of 256 to 2,048 dimensions as float, int8 or binary. Cohere recommends 1,024-dimensional int8 vectors, and notes that a 256-dimensional binary vector takes 32 bytes against 8 KB for a 2,048-dimensional float32 one.

Much of the comparison uses RCP-nDCG@10, Cohere's new metric that scores retrieved documents against query-specific relevance criteria rather than a fixed set of labels; Embed 5 is the first family evaluated with it. On ViDoRe V3 (visually rich documents from eight enterprise domains), Embed 5 Pro averages 85.8, 8.8 points above Embed 4, against 83.7 for Voyage 4 Large, 83.2 for Gemini Embedding 2 and 75.5 for OpenAI text-embedding-3-large; Fast averages 84.5. Pro ranks first on FinanceBench (80.1), FinQA (90.0) and ViDoRe V3 Finance (85.0), with Fast second on each, and leads Cohere's parsed-PDF suite with 84.8 against 83.6 for Voyage 4 Large and 80.8 for Gemini Embedding 2. Across German, French, Spanish, Italian and Russian, Pro averages 77 against 76 for Voyage 4 Large and 73 for Gemini Embedding 2, but in Cohere's own table Gemini Embedding 2 scores higher on Japanese, Korean and Arabic. On ViDoRe V3, Fast leads compact models such as Voyage 4 Nano (by almost seven points) and Microsoft's Harrier 0.6B (by ten or more). All scores are Cohere's.

Weights are closed. Embed 5 is generally available on the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker and inside Cohere's North platform, and both tiers can be deployed privately in a VPC or on premises with vLLM. Pricing per million text tokens is $0.12 for Pro and $0.08 for Fast, and $0.40 per million image tokens for both. Cohere gives no parameter counts, saying only that Fast is roughly half the size of Qwen3-VL-Embedding-2B.