Cohere launches Parse 5, a vision language model for structured document conversion

Cohere Introduces Parse 5 for Document Processing

Cohere has launched Parse 5, a new vision language model aimed at enterprises needing to convert various document types into structured data. The model is specifically designed to handle complex layouts found in PDFs, slides, and scanned images, transforming them into Markdown format while preserving structural elements like tables and charts.

Focus on Cost-Performance for Enterprise Scale

Parse 5 is positioned by Cohere for its price-to-performance ratio rather than achieving the highest accuracy scores. Cohere's own benchmarks indicate that Parse 5 trails larger models like GPT-5.5, Opus 4.8, and Gemini 3.5 Flash in accuracy. However, its pricing of $1.50 per 1,000 pages through its API, with Model Vault for higher-volume deployments, targets the need for cost-effective solutions at enterprise scale.

Addressing Structural Challenges in Document Parsing

Nils Reimers, VP of AI Search at Cohere, stated that preserving structure and meaning is the difficult aspect of document parsing, as enterprise documents often contain mixed content like tables, diagrams, and charts that affect data interpretation. Many existing tools either fail to retain this structure or produce inaccurate content, even with advanced models.

Single-Pass Architecture and Capabilities

Parse 5 utilizes a single-pass architecture, processing a page as an image through one vision-language model pass to return structured Markdown. This approach consolidates the OCR and model pipeline typically handled in separate steps. The model is built on Cohere Labs' North-Micro-Vision-Instruct architecture, featuring a 2.3-billion-parameter size, an 8,192-token context window, and a 4.6-gigabyte footprint.

It accepts base64-encoded images from PDFs, PowerPoints, or JPEGs, returning Markdown in reading order. Tables are rendered as HTML, and image descriptions along with bounding box coordinates for tables and images are provided. Parse 5 supports stable accuracy for Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish, with zero-shot support for other languages. Output modes include a default Markdown string per page and a blocks mode that returns typed elements with HTML, bounding box, and description for each table.