Cohere launches Parse 5: a 2.3-billion-parameter vision-language model that turns documents into Markdown

Cohere has released Parse version 5 (parse-v5.0), a 2.3-billion-parameter vision-language model built on Cohere Labs' North-Micro-Vision-Instruct architecture. The model accepts PDF pages, PPT presentations or JPEG images encoded as base64 URIs and returns Markdown in a single pass — reading-order text, tables as HTML, lists, form key-value pairs, image descriptions and bounding-box coordinates — without a separate OCR step. The context window is 8,192 tokens and the total weight is roughly 4.6 GB, making it deployable in resource-constrained environments.

Partial benchmark and competitive picture

The company reports a ParseBench score of 79.2, but that figure is an average of only three of the benchmark's five dimensions — tables, content fidelity and semantic layout — omitting charts and visual grounding, the two dimensions where most parsers tend to fail. On the same partial calculation, Mistral OCR 4 scores 74.5, Azure Document Intelligence 74.3 and Databricks AI Parse 72.4. On the full public leaderboard, however, those three competitors cluster around 59–60, while LlamaParse Agentic leads with 84.88. Cohere does not appear on the public board, so the 79.2 is a self-reported score on a subset, not a leaderboard position.

Deployment, pricing and business model

The model is already available in production through Cohere's Parse API, Microsoft Foundry, AWS SageMaker and Model Vault for single-tenant deployments, with no waitlist and no research licence. Pricing is $1.50 per thousand pages, and Cohere positions the product on price-performance rather than peak accuracy. The company says the economics become meaningful only above roughly 100 thousand pages per month, targeting mid-market teams that already run a RAG stack and large enterprises with data-residency or air-gap requirements.

Output modes and target industries

Two output modes are practically relevant: the default returns a Markdown string per page, while setting output_format="blocks" returns typed blocks — a table block carries its HTML, bounding box and description. The second mode enables citation-level traceability, a critical requirement in intelligent document processing (IDP) pipelines. Nine languages are designated as stable — Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish — with zero-shot support for others at lower accuracy. Cohere is aiming at document-heavy sectors: financial services, insurance, healthcare and life sciences, public sector, telecom, energy and manufacturing.

What this means for RAG pipelines

For teams building ingestion pipelines for RAG, Parse offers a lightweight alternative that skips a separate OCR chain and returns structure rich enough for source citation. The gap between the reported three-dimension score and the significantly lower results on the full leaderboard is a reminder to run an evaluation on the organisation's representative documents before committing at high volume. In the absence of a public leaderboard presence, the 79.2 is a reasonable reference point for testing — nothing more.