The Pulse
Cohere Launches Embed 5 With Pro and Fast Search Models
Cohere launched Embed 5 on September 30, pairing a high-quality model with a faster, lower-cost option for enterprise search. The models handle text, images and mixed document pages, and share an embedding space so companies can index with
AI.info Team · Written with AI ·
Enterprise search systems often flatten PDFs into text before indexing them, losing table relationships, reading order and information embedded in charts. Cohere says its new Embed 5 models can instead represent text, images and mixed text-and-image pages in a single vector, though the performance figures it cites come from its own evaluations. The company announced the models on September 30, offering one tier for retrieval quality and another for speed and cost.
One index, two models
Embed 5 comes in Pro and Fast versions. Cohere positions Pro for high-quality offline indexing and complex corpora, while Fast targets interactive search, agent workflows and other high-volume requests. The two models share an embedding space, allowing teams to index documents with Pro and query that index with Fast rather than rebuilding it.
“Today, we're releasing Embed 5, a new family of embeddings models at the frontier of high-quality enterprise retrieval.” Cohere Team
Both models accept text, images and fused text-image inputs, support more than 100 languages and have a 128,000-token context window. Customers can choose output dimensions from 256 to 2,048 and use float, int8 or binary formats. Cohere lists pricing at $0.12 per million text tokens for Pro and $0.08 for Fast; image input costs $0.40 per million tokens for either model.
PDF structure is the test
Cohere’s central pitch is that enterprise documents contain useful information that text extraction can discard: a table’s rows and columns, a multi-column page’s reading order, or a chart’s visual content. In the company’s reported tests on ViDoRe V3, Embed 5 Pro scored an average 85.8 across eight visually rich document domains, compared with 77.0 for Embed 4. Cohere reported 84.5 for Fast, 83.7 for Voyage 4 Large and 83.2 for Google’s Gemini Embedding 2.
On a separate parsed-document suite covering material such as service documentation, corporate reports, SEC filings and privacy policies, Cohere reported an average of 84.8 for Pro, 83.4 for Fast and 83.6 for Voyage 4 Large. The company also reported that Pro averaged 82.3 on five fused text-image datasets, against 81.2 for Fast and 61.3 for Gemini Embedding 2. Those are company-published results, not an independent comparison; Cohere says the ViDoRe V3 documents were curated by the benchmark’s authors, while some other tests include internal data.
Benchmarks measure only part of retrieval
Cohere introduces Embed 5 alongside RCP-nDCG@10, a retrieval evaluation method it says assesses how well models order documents against query-specific relevance criteria. The company’s own footnote sets a boundary on that result: the method reorders a fixed candidate set in a two-stage retrieval setup, so its scores reflect reranking quality rather than first-stage retrieval. The distinction matters because strong ranking among supplied candidates does not, by itself, show how well a system finds those candidates in a large index.
Pro and Fast were tested in combinations across 40 development datasets, according to Cohere. The company says using Fast for queries against a Pro-indexed corpus produced an average 1.6% quality loss relative to the Pro-on-Pro baseline, while querying a Fast-indexed corpus with Pro produced a 2.7% loss. Cohere recommends Pro for indexing and Fast for queries, a split intended to preserve much of the reported retrieval quality while lowering query costs.
Available through three services
Embed 5 is available through Cohere’s API and Model Vault, Microsoft Foundry and Amazon SageMaker. Cohere also says both models can be served with vLLM for private deployments in a customer’s own virtual private cloud or on-premises, and lists integrations with tools and vector databases including Elasticsearch, Pinecone, Weaviate and OpenSearch.
The release gives enterprise teams a new way to trade retrieval quality against query speed without maintaining separate embedding spaces. Whether that arrangement improves search on a company’s own files will depend on its documents, candidate-generation pipeline and evaluation method—the elements the launch benchmarks cannot settle for every deployment.