Cohere Parse

// Overview

Cohere Parse, invoked as `parse-v5.0`, is a document-parsing vision-language model Cohere published on 27 August 2026. It takes PDFs, PowerPoint decks and base64-encoded JPEG images and returns a single Markdown document: text in reading order, lists, tables rendered as HTML, bounding-box coordinates for visual elements, and descriptions of embedded images. The point is to collapse what is usually a chain of separate stages — OCR, a layout model, a table extractor — into one call whose output a downstream model or retrieval pipeline can consume directly.

The model is deliberately small. Cohere's documentation lists 2.3B parameters, an 8,192-token context window and a footprint of roughly 4.6GB, on a proprietary architecture derived from Cohere Labs' North-Micro-Vision-Instruct. That size is the product argument: Cohere reports throughput of 4.5 pages per second, rising to 36 pages per second on an eight-H100 node, and prices the hosted API at $1.50 per 1,000 pages. Nine languages are supported — Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish.

On Cohere's own ParseBench comparison the model averages 79.2 across tables, content faithfulness and semantic formatting. That places it ahead of the dedicated parsing stacks it is aimed at — LlamaParse (Cost Effective) at 78.3, Chandra OCR 2 at 77.7, Mistral OCR 4 at 74.5, Databricks AI Parse at 72.4, Azure Document Intelligence at 69.3, Google Document AI at 57.3 and AWS Textract at 53.3 — and behind the general-purpose frontier models in the same table, GPT-5.5 at 84.4, Opus 4.8 at 84.3 and Gemini 3.5 Flash at 81.8. Its strongest columns are tables (87.0) and content faithfulness (86.6); semantic formatting (64.0) is its weakest.

Besides the hosted Cohere API, Parse is offered through Cohere's Model Vault for private deployment and on Microsoft Foundry and AWS SageMaker. Cohere reports that Model Vault runs 23% below the API price at 50% GPU utilisation and up to 61% below at full hourly utilisation, which is the trade being pitched to teams with a steady, high-volume document backlog rather than bursty traffic.

// Benchmarks

This model's scores

1. ParseBench — tables87
2. ParseBench — content faithfulness86.6
3. ParseBench — average79.2
4. ParseBench — semantic formatting64

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Flat per-page price on the hosted Cohere API — pages parsed, not tokens, and no separate output charge. Cohere reports that a Model Vault private deployment costs 23% less than the API at 50% GPU utilisation and up to 61% less at full hourly utilisation.

Pricing source ↗

// Strengths

• One call replaces an OCR-plus-layout-plus-table-extraction pipeline: Markdown, HTML tables, bounding boxes and image descriptions come back together
• 2.3B parameters and a ~4.6GB footprint, at 4.5 pages per second (36 pages per second on an 8×H100 node)
• $1.50 per 1,000 pages on the hosted API, with Model Vault cheaper still at sustained GPU utilisation
• 87.0 on ParseBench tables and 86.6 on content faithfulness — the two columns enterprise document work leans on hardest
• Deployable privately through Model Vault, Microsoft Foundry and AWS SageMaker, not only as a hosted endpoint

// Best for

• Turning contract, invoice and report archives into clean Markdown for a retrieval-augmented pipeline
• Table-heavy extraction — financial statements, lab results, price lists — where HTML table fidelity matters more than prose
• Bulk backfills of scanned document stores, where per-page cost and pages-per-second decide whether the job is affordable
• Regulated or air-gapped deployments that need document parsing to run inside the customer's own environment

// How to access

// Cohere Parse in the news

1. 2026-08-27MAJORCohere Parse — a 2.3B document model that turns PDFs into clean Markdown

From the AI/TLDR release feed — every item is source-verified when it ships.