Cohere has released Parse (parse-v5.0), a document parsing model aimed at high-volume enterprise ingestion. It is a 2.3B-parameter vision language model with an 8,192-token context window and a ~4.6GB footprint, built on Cohere Labs’ North-Micro-Vision-Instruct architecture. Parse takes a PDF, PPT or JPEG page as a base64-encoded data URI and returns Markdown containing text in reading order, tables rendered as HTML, lists, form key-value pairs, image descriptions and bounding box coordinates. There is no separate OCR stage in front of it. Cohere prices the Parse API at $1.50 per 1,000 pages and positions the model on price-performance rather than peak accuracy — a claim the company supports with a self-reported ParseBench score of 79.2 that, as we detail below, measures three of that benchmark’s five dimensions.
Is it deployable?
Yes, in production. Parse is generally available through the Cohere Parse API, Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault. There is no waitlist and no research license.
- Which companies: Mid-market teams that already run a RAG stack can start on metered API calls with a free trial key. Large enterprises with residency or air-gap requirements go straight to Model Vault or private deployment. Seed-stage startups can use it, but the economics only start to matter above roughly 100K pages a month.
- Which industries: Cohere targets financial services, insurance, healthcare and life sciences, public sector, telecom, energy and manufacturing — the document-heavy verticals where scanned forms and dense tables are the norm.
- Applications: RAG ingestion, intelligent document processing, claims and invoice pipelines, contract and filing search, and giving document context to agents.
What is Parse?
Parse is a 2.3B-parameter vision language model built on Cohere Labs’ North-Micro-Vision-Instruct architecture, with an 8,192-token context window and a ~4.6GB footprint. It accepts PDF, PPT and JPEG pages as base64-encoded data URIs and returns Markdown containing document text, lists, tables rendered as HTML, bounding box coordinates and image descriptions.
There is no separate OCR stage in front of it. The model recovers text and reading order, tables, lists, forms and key-value pairs, images and captions, and the locations of page boundaries and visual elements in one pass. Nine languages are listed as stable — Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish — with zero-shot support elsewhere at lower accuracy.
Two output modes matter in practice. The default returns a Markdown string per page. Setting output_format="blocks" returns typed blocks, where a table block carries its HTML, its bounding box and a description. That second mode is what makes citation-level traceability possible.
‘;
}).join(”);
}
var drawn=false;
function drawBars(){
if(drawn)return;drawn=true;
rows(document.getElementById(‘setA’),A);
rows(document.getElementById(‘setB’),B);
setTimeout(function(){
[].slice.call(document.querySelectorAll(‘.fill’)).forEach(function(f){f.style.width=f.dataset.w+’%’});
size();
},60);
}
function size(){
try{parent.postMessage({mtpHeight:document.body.offsetHeight+40},’*’)}catch(e){}
}
window.addEventListener(‘load’,size);
window.addEventListener(‘resize’,size);
setTimeout(size,300);setTimeout(size,900);
})();
