Cohere Parse Document Extraction Model Now Available in Microsoft Foundry

Cohere Parse, the company's vision language model for extracting structured data from complex documents, is now available in Microsoft Foundry, joining Cohere's existing Embed and Rerank models in the Azure-hosted platform. The model converts multimodal files—tables, forms, diagrams, embedded images—into clean Markdown rather than a flat character stream, according to the vendor.

Parse runs as a direct-from-Azure offering that deploys into the customer's own Foundry environment. Inference stays within the tenant boundary and inherits existing network, identity, and data governance controls. For regulated industries with residency or isolation requirements, that removes a common objection to externally hosted document-processing APIs.

Cohere says Parse returns bounding boxes for visual elements, which supports retrieval, grounding, and citation in downstream applications. The coordinate data allows downstream systems to point back at the exact region of a source document that produced a given piece of extracted content.

Cohere reports throughput of 4.5 pages per second per GPU, or roughly 36 pages per second on an 8× H100 node, served with vLLM. These are the vendor's own measurements; real-world numbers will depend on document complexity and hardware configuration. Pricing is set at $1.50 per thousand pages for Cohere-Parse-v5.

Cohere says Parse was trained on business documents across finance, insurance, and scientific research, and supports nine major languages. That domain focus may improve accuracy on the document types it was trained on, but it also narrows the model's comfort zone. Teams processing other document types, such as legal filings or engineering specs, will want to validate output against their own corpus before assuming the model's enterprise claims transfer.

The model is available through the Azure AI model catalog. For prospective users, the question is whether the nine-language, domain-focused training set matches the documents they actually process, and whether the per-page cost and throughput hold up against their own pipeline's volume.