Cohere Parse v5.0 Turns Complex Documents Into AI-Ready Markdown

Cohere releases a dedicated multimodal document-parsing model

Cohere released Parse, model ID parse-v5.0, on August 27, 2026. The 2.3-billion-parameter multimodal model converts complex enterprise documents into structured Markdown that can feed retrieval, search and agent workflows.

Rather than treating a document as plain OCR text, Parse extracts text in reading order, tables, lists, forms and key-value pairs, images and captions, page boundaries, and the locations of visual elements. Its responses can include Markdown/HTML content, HTML-formatted tables, bounding boxes and generated image descriptions.

Cohere lists an 8,192-token context window and an approximately 4.6GB model size. The currently documented input formats are PDF, PowerPoint and base64-encoded JPEG images.

The main use case is document preparation for RAG and agents

Cohere positions Parse as an ingestion layer for high-volume enterprise document workloads. A parsed document can be passed into search and retrieval systems, used to build a RAG index, or supplied as grounded context to an AI agent.

Nine input languages are listed as stable: Arabic, English, French, German, Japanese, Korean, Italian, Portuguese and Spanish. The model can attempt additional languages zero-shot, but Cohere warns that accuracy may be lower.

The model is available through Cohere's Parse API, Microsoft Foundry and Amazon SageMaker. Cohere also offers single-tenant deployment through Model Vault. The model catalog lists parse-v5.0 as Live, so this is a released service rather than a preview announcement.

Important limitations remain for structured-document workflows

Cohere documents several boundaries that matter in production. Parse does not return confidence scores for extracted content, does not identify headers, footers or font hierarchy, and its primary document representation is Markdown rather than arbitrary structured JSON.

Those limits mean teams handling regulated or high-stakes documents should still validate extraction quality, especially for tables, dense forms, unusual layouts and languages outside Cohere's stable list. Bounding boxes and block output can help downstream verification, but they are not a substitute for application-specific accuracy testing.

The release is notable because it gives Cohere a dedicated, compact document-understanding layer that can sit before Embed, Rerank, Compass or external agent systems instead of forcing a general-purpose chat model to perform every parsing step.