Cohere Releases Parse 5 2.3B Vision Language Model
Cohere has launched Parse 5 (parse-v5.0), a 2.3B-parameter vision language model that converts enterprise documents into Markdown in a single pass.
Stock photo for illustration only, not from the actual event
- Parse 5 is a 2.3-billion parameter vision language model developed by Cohere.
- It converts PDF, PPT, and JPEG documents directly into Markdown in one pass.
- Achieved a ParseBench score of 79.2, outperforming several industry peers.
- Generally available via Cohere Parse API, Microsoft Foundry, and AWS SageMaker.
Cohere has introduced Parse 5 (parse-v5.0), a 2.3-billion-parameter vision language model built on Cohere Labs’ North-Micro-Vision-Instruct architecture. The model is engineered to transform enterprise documents into clean Markdown formats, accepting PDF, PPT, and JPEG pages provided as base64-encoded data URIs. It features an 8,192-token context window and a footprint of approximately 4.6 gigabytes.
A key architectural advantage is the absence of a separate preliminary OCR stage. Instead, the model recovers textual content, reading orders, tables, lists, forms, key-value pairs, images with captions, page boundaries, and visual element locations in a single pass. It natively supports nine stable languages—Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish—while offering zero-shot capabilities for other languages at a reduced accuracy.
Stock photo for illustration only, not from the actual event
In operational environments, the model supports two distinct output modes. The default mode returns a single Markdown string per page, whereas setting output_format="blocks" returns structured typed blocks, such as table blocks containing associated HTML code, bounding box coordinates, and descriptions. This secondary mode enables precise citation-level traceability.
Regarding benchmark performance, Cohere reported a ParseBench score of 79.2 for Parse, averaged across table extraction, content faithfulness, and semantic formatting dimensions. This outperforms Mistral OCR 4 at 74.5, Azure Document Intelligence at 74.3, and Databricks AI Parse at 72.4. ParseBench itself is a LlamaIndex benchmark comprising roughly 2,078 human-verified enterprise pages evaluated across five criteria. Cohere’s reported figure averages three of those dimensions while excluding charts and visual grounding—areas where traditional document parsers typically struggle.
Eliminating the traditional decoupled OCR pipeline in favor of end-to-end multi-modal document understanding represents a significant architectural shift in handling enterprise document workflows. However, practitioners should carefully analyze benchmark metric subsets against their specific document layouts and types to determine real-world utility.
Parse is currently available for production use through the Cohere Parse API, Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault deployments, requiring no waitlist or specialized research license. API usage is priced at $1.50 per 1,000 pages. On Model Vault, a Medium instance is priced at $4.00 per hour or $2,500 per month, while an XL instance costs $7.00 per hour or $4,300 per month, allowing enterprises to optimize costs based on their monthly processing volume and data residency requirements.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment