Cohere launches Parse 5 to turn documents into Markdown

Cohere has released Parse, version parse-v5.0, a 2.3-billion-parameter vision-language model built to convert enterprise documents into Markdown at scale. According to trade outlet MarkTechPost, which first reported the release on August 27, and corroborated separately by South Korean outlet AI Times, the system takes a PDF, PowerPoint or JPEG page encoded as a base64 data URI and returns the text in its original reading order, tables rendered as HTML, lists, form key-value pairs, image descriptions, and the bounding-box coordinates of every element on the page.

No separate OCR stage

Unlike classic document-digitization pipelines, which run a document through a separate optical-character-recognition engine first, Parse does that work in a single pass: the vision-language model itself — built on Cohere Labs' North-Micro-Vision-Instruct architecture — recovers text, page layout, tables, forms and images all at once. It has an 8,192-token context window and a roughly 4.6-gigabyte footprint. Cohere offers two output modes: a default that returns a plain Markdown string, and a 'blocks' mode that returns typed blocks — each carrying its HTML, its exact position on the page and a description — designed so a retrieval system can cite precisely which part of a document a given fact came from. According to AI Times, the model stably supports nine languages, including English, Korean, French, German, Japanese and Spanish, and processes 36 pages per second — 2,160 per minute — running on eight H100 GPU nodes, a speed the outlet puts at roughly 2.2 times faster than the open-source document parsers it competes with.

A figure Cohere chose itself

Cohere backs its price-over-peak-accuracy pitch with a ParseBench score of 79.2. ParseBench is LlamaIndex's benchmark of roughly 2,078 human-verified enterprise pages, scored across five dimensions: tables, charts, content faithfulness, semantic formatting and visual grounding. That 79.2 figure, however, averages only three of those five dimensions — it drops charts and visual grounding, the two dimensions where most document parsers struggle most. On that same three-dimension basis, Parse edges out Mistral OCR 4 (74.5), Azure Document Intelligence (74.3) and Databricks AI Parse (72.4); AI Times adds AWS Textract (53.3) and Google Document AI (57.3) to the comparison, and notes that only frontier language models — GPT-5.5, Opus 4.8 and Gemini 3.5 Flash — score higher on that test. Measured against the full five-dimension public leaderboard, the same rivals fall well below that: Mistral OCR 4 and Databricks AI Parse manage only 60.68 points, and Azure Document Intelligence (Layout) 59.64 — and Cohere Parse does not yet appear on that full leaderboard, where LlamaParse Agentic leads with 84.88.

• API pricing: $1.50 per 1,000 pages.
• Model Vault, Medium instance: $4.00 an hour, or $2,500 a month.
• Model Vault, XL instance: $7.00 an hour, or $4,300 a month.
• Break-even point versus metered API pricing: roughly 1.67 million pages a month for Medium, and 2.87 million for XL.
• Extra savings from deploying on Model Vault or on-premises, per AI Times: up to 61% versus the standard per-page rate.

Who it's built for

Parse is already available in production, with no waitlist and no research license, through the Cohere API, Microsoft Foundry, AWS SageMaker and Model Vault, its single-tenant dedicated-instance option. Cohere is targeting paperwork-heavy industries — banking and financial services, insurance, healthcare and life sciences, the public sector, telecom, energy and manufacturing — for tasks such as feeding retrieval-augmented generation (RAG) pipelines, processing claims and invoices, searching contracts and filings, or giving document context to AI agents. Teams that already run a RAG stack can start with metered API calls and a free trial key, while large enterprises with data-residency or air-gap requirements go straight to Model Vault or a private deployment; the economics of a dedicated instance, per MarkTechPost, only start to make sense above roughly 100,000 pages a month.

For a company that processes large volumes of invoices, claims or contracts — in insurance, banking or the public sector — Parse offers a known-cost alternative to tools like Azure Document Intelligence or Amazon Textract, with English among the nine languages it supports natively. The performance figure Cohere is promoting is still worth reading cautiously: measured against the full five-dimension public benchmark, neither Parse nor its main commercial rivals come close yet to the specialized systems leading that leaderboard. Before moving a critical workflow over, testing the model on a company's own documents remains the only check that actually settles the question.