Cohere has released Embed 5, a new family of frontier embedding models available in Pro and Fast tiers on the Cohere API, Microsoft Foundry, and Amazon SageMaker. The two tiers share a single embedding space, allowing teams to index with Pro for maximum quality and query with Fast for lower latency and cost without rebuilding the index.
- Embed 5 Pro achieves an average score of 85.8 on ViDoRe V3, outperforming Voyage 4 Large (83.7) and Gemini Embedding 2 (83.2), while ranking first on FinanceBench, FinQA, and ViDoRe V3 Finance.
- Embed 5 Fast offers a third of the cost of Pro with 2.4x higher document throughput, averaging 84.7 on ViDoRe V3 and leading compact models like Voyage 4 Nano.
- Both models support multimodal inputs, multilingual coverage across 100+ languages, and Matryoshka representation learning to reduce vector storage size by up to 256x using int8 quantization.
The release enables enterprises to improve answer quality and user experience in search and RAG workflows while keeping downstream inference costs under control.