Cohere Releases Embed 5: Pro and Fast Models Share Embedding Space for Enterprise Search

Cohere has released Embed 5, a new family of embedding models comprising Pro and Fast tiers. The core architectural innovation is a shared embedding space, allowing users to index documents with the high-quality Pro model and query with the low-latency Fast model without re-indexing. This hybrid approach reduces costs by up to one-third for query operations while maintaining near-baseline search quality.

Performance and Benchmarks

Embed 5 Pro demonstrates significant improvements in enterprise document retrieval. On the ViDoRe V3 benchmark for visually rich documents, Pro scores an average of 85.8, an 8.8-point increase over Embed 4. It surpasses competitors including Voyage 4 Large (83.7), Gemini Embedding 2 (83.2), and OpenAI text-embedding-3-large (75.5). In financial domains, Pro leads on FinanceBench (80.1) and FinQA (90.0), outperforming the next best non-Cohere model by an average of 3.3 points.

The Fast tier is optimized for latency-sensitive workloads, offering 2.4x higher throughput than Pro. It achieves a ViDoRe V3 score of 84.5, beating smaller competitors like Jina v5 Small (74.5) and Qwen3-VL-Embedding-2B (64.2) by substantial margins.

Multilingual and Multimodal Capabilities

Embed 5 supports over 100 languages with a 128K token context window. In European languages, Pro averages 77 points, with notable gains in Persian (+13), Telugu (+12), and Hindi (+12) compared to Embed 4. For multimodal tasks, Pro handles fused text-image inputs with an average score of 82.3 across five datasets, significantly outperforming Fast (81.2) and Gemini Embedding 2 (61.3).

Deployment and Optimization

The models are available via Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker. Cohere introduces RCP-nDCG@10, a new evaluation metric that better reflects corpus-wide performance by accounting for query-specific relevance criteria. To optimize storage, Embed 5 supports Matryoshka embeddings and quantization formats (float, int8, binary). Using 256-dimensional binary vectors can reduce storage requirements by up to 256x compared to 2,048-dimensional float32 vectors, dropping raw vector storage for 100 million chunks from ~819 GB to ~3.2 GB.