Cohere Introduces Embed 5: Dual-Tier Models for Enterprise AI
Cohere, a prominent player in artificial intelligence, has officially unveiled its new Embed 5 model family. This release specifically targets significant advancements in enterprise search, Retrieval Augmented Generation (RAG), and agentic retrieval applications. The Embed 5 family is launched in two distinct versions: Embed 5 Pro, engineered for maximum retrieval precision, and Embed 5 Fast, optimized for low latency and cost-efficiency during live query operations. Both models accept a diverse range of inputs including text, images, and combined text-plus-image data, supporting over 100 languages with a substantial context window of up to 128,000 tokens.
A pivotal design element of this new family is the unified embedding space shared by both Pro and Fast tiers. This allows organizations the flexibility to index their data using one model and query it with the other, streamlining workflows while balancing performance and cost.
Deployment Accessibility and Technical Specifications
The Embed 5 models are immediately available for broad deployment. Users can access both tiers through the Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker. For environments requiring heightened security or specific infrastructure, private VPC or on-premise serving options are facilitated via vLLM.
According to Cohere’s model documentation, the API identifiers are embed-v5.0-pro and embed-v5.0-fast. These models are capable of generating embeddings in various dimensions, with 2048 as the default, alongside options for 1536, 1024, 768, 512, or 256. Output types include float, int8, or binary formats. Regarding cost, Embed 5 Pro is priced at $0.12 per million text tokens, while Embed 5 Fast costs $0.08 per million text tokens. Image inputs across both tiers are uniformly priced at $0.40 per million tokens.
Multimodal Capabilities and Strategic Usage
Embed 5 offers sophisticated multimodal embedding, capable of directly processing page images. Furthermore, it can fuse an image with its associated metadata into a single vector. This capability is particularly crucial for handling complex visual documents such as scanned pages, slide presentations, schematics, and charts, where traditional text extraction methods might lose valuable contextual information.
Cohere’s recommended strategy leverages the shared embedding space: indexing a corpus with the Embed 5 Pro model and subsequently querying it with Embed 5 Fast. Internal testing, covering numerous corpus and query combinations across 40 development datasets, revealed that a Pro-indexed corpus queried by Fast achieved a retrieval quality score of 98.4 (normalized against a 100-point Pro-plus-Pro baseline), while an all-Fast setup scored 96.6. This approach is especially beneficial for agentic workloads where query latency is a critical factor, as Embed 5 Fast processes documents at a reported rate of 377.3 per second, significantly faster than Pro's 159.7 documents per second.
Performance Benchmarks and Storage Efficiency
On the ViDoRe V3 benchmark, Embed 5 Pro achieved an average score of 85.8, marking an 8.8-point improvement over its predecessor, Embed 4, with Embed 5 Fast following closely at 84.5. These scores position Cohere's latest models favorably against competitors, outperforming Voyage 4 Large (83.7), Gemini Embedding 2 (83.2), and OpenAI's text-embedding-3-large (75.5). In Cohere’s parsed-PDF suite, Pro led with 84.8 compared to Voyage 4 Large’s 83.6. The models demonstrated particular strength in financial applications, with Pro securing the top rank on FinanceBench (80.1), FinQA (90.0), and ViDoRe V3 Finance (85.0).
However, multilingual performance showcased varied results. While Pro led in the European-language average with 77, Gemini Embedding 2 demonstrated superior performance over Pro in nine out of ten additional languages tested by Cohere, including Japanese, Arabic, Hindi, and Telugu. It is important to note that many of these reported figures utilize RCP-nDCG@10, a new metric developed by Cohere that focuses on reranking quality. While the evaluation code is public, independent validation remains pending.
Storage efficiency is another highlight, thanks to Matryoshka representation learning and lower-precision outputs. For instance, a 2048-dimension float32 vector, typically requiring 8 KB, can be drastically reduced. A 1024-dimension int8 vector needs only 1 KB, and a 256-dimension binary vector uses just 32 bytes. Scaling this across 100 million chunks, raw storage requirements could decrease from approximately 819 GB to a mere 3.2 GB. Cohere suggests using 1024-dimension int8 as the default, citing its ability to retain near full-precision quality.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
