Cohere Launches Parse 5: A 2.3B Vision Language Model for Enterprise Document Processing

Cohere has released Parse 5, a 2.3-billion-parameter vision language model built on the North-Micro-Vision-Instruct architecture. [3] The model features an 8,192-token context window and an approximate 4.6GB footprint. [3] Without using a separate OCR stage, Parse accepts PDF, PPT, or JPEG pages as base64-encoded data URIs and returns Markdown files. [3] The returned Markdown includes text in reading order, tables rendered as HTML, form key-value pairs, and image descriptions. [3] The model is capable of processing documents across nine major world languages. [2]

Parse is priced at $1.50 per 1,000 pages through the Cohere API. [2][3] The model targets document-heavy industries by focusing on price-performance. [2] The model is generally available in production via Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault deployments. [3] On a three-dimension average of ParseBench, Cohere reports that Parse scores 79.2. [2][3]

Cohere is positioning Parse 5 to compete on price-performance for high-volume enterprise ingestion rather than peak accuracy. By eliminating a discrete OCR step, the 2.3B-parameter model aims to streamline RAG data preparation pipelines and reduce architectural complexity. The company's self-reported ParseBench score of 79.2 places it ahead of Mistral OCR 4 (74.5) and Databricks AI Parse (72.4), as well as LlamaParse's Cost Effective tier (78.3). This pricing and architecture combination suggests a strong push into document-heavy verticals like financial services and healthcare where throughput economics matter most. What the sources don't address: How the model's extraction accuracy degrades on highly corrupted or handwritten legacy documents compared to dedicated, multi-stage OCR pipelines.