Cohere Introduces Parse, a Vision Language Model for Document Processing

Cohere Parse is a vision language model designed to process large volumes of enterprise documents into structured, machine-readable data. [2] The model detects visual elements such as tables and embedded images and outputs clean Markdown files. [2] It supports nine major world languages and can be used for document indexing, RAG, and agentic retrieval. [2]

Customers can access Parse via the Cohere API for $1.50 per 1,000 pages or deploy it on their own infrastructure. [2]

On the ParseBench evaluation, which measures agent-suitable parsing performance, Cohere Parse scored 79.2. [2] This score compares to 74.5 for Mistral OCR 4, 72.4 for Databricks AI Parse, and 78.3 for LlamaParse’s Cost Effective offering. [2] The model also offers an over 20-point improvement compared to hyperscaler document intelligence solutions. [2]

Cohere is explicitly positioning Parse as a cost-effective alternative for enterprise document processing workloads. By citing scores against Mistral OCR 4, Databricks AI Parse, and LlamaParse, Cohere highlights a competitive landscape where parsing quality and pricing are critical differentiators for enterprise adoption. The inclusion of Markdown output specifically targets RAG and agentic retrieval pipelines, which rely heavily on clean text structuring. What the sources don't address: how the model's processing speed and latency compare to the listed competitor offerings during peak production workloads.