Cohere is betting that a small model can replace the traditional OCR-plus-LLM pipeline for enterprise documents. Parse 5, its newest release, is a 2.3B-parameter vision language model with an 8,192-token context window and a footprint of about 4.6GB, built on Cohere Labs’ North-Micro-Vision-Instruct architecture.
Give it a page and it returns structured output: text in reading order, tables as HTML, lists, form key-value pairs, image descriptions, and bounding box coordinates. PDF, PPT, and JPEG pages go in as base64 data URIs, and there is no separate OCR stage. Nine languages are listed as stable, including Arabic, English, French, German, Italian, and Japanese.
Cohere’s pricing is aggressive, at $1.50 per thousand pages, and the company leans on price-performance over peak accuracy, pointing to a self-reported ParseBench score of 79.2 as evidence. There is no waitlist and no research license. The model is generally available through the Cohere Parse API, Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault.
For teams feeding invoices, contracts, and slide decks into agent workflows, one small model now handles what used to take two systems and a pile of integration work.
