Cohere ships Parse, a document-reading model
Archive item — written before sources were shown.
Cohere's new Parse model turns messy enterprise PDFs into structured Markdown, beating Textract and Document AI on its own benchmark.
Cohere launched Parse on August 27, a 2.3-billion-parameter vision-language model built specifically to convert enterprise documents, contracts, invoices, forms, and scanned reports, into structured, machine-readable Markdown. It goes beyond plain OCR: Parse detects tables, forms, diagrams, and images, and returns bounding boxes for visual elements so document structure survives the conversion, which matters for retrieval-augmented generation and other agentic workflows that need to cite a specific cell or clause.
On Cohere’s own ParseBench evaluation, Parse scored 79.2 overall, ahead of specialized document tools like Mistral OCR 4 (74.5), Databricks AI Parse (72.4), and Azure Document Intelligence (69.3), with a table-extraction score of 87.0. It trails general frontier multimodal models on raw accuracy, GPT-5.5 and Opus 4.8 both score higher on the same benchmark, but at a fraction of the cost: the model processes documents at 4.5 pages per second on a single call, scaling to 36 pages per second across an 8-GPU node, and Cohere prices it at $1.50 per 1,000 pages. It is available through Cohere’s own API, a single-tenant Model Vault deployment, or private cloud and on-premises installs, plus listings on Microsoft Foundry and AWS SageMaker.
The pitch
Document parsing is a boring but expensive part of most agentic pipelines, and Cohere is positioning Parse on price-to-performance rather than best-in-class accuracy: cheap and fast enough to run at the volume real document pipelines need, without paying frontier-model rates for what is fundamentally a structured-extraction task. It beats the incumbent cloud OCR services on Cohere’s own numbers while undercutting a general-purpose multimodal model on cost per page. For finance, insurance, and legal teams processing volume paperwork, this is the kind of infrastructure swap worth a real side-by-side test against whatever Cohere’s earlier North Automations or existing pipeline already handles, since the savings compound fast at scale.
- 01Introducing Parse: Enterprise document intelligence at scalecohere.com · primary
- 02Cohere Parse 5 loses the benchmark on points. It wins on cost per page.venturebeat.com · independent reporting
