Cohere Parse: A 2.3B-Parameter Document Structuring Model
Cohere released Parse on August 27, 2026. This 2.3B-parameter multimodal model converts complex documents into structured Markdown and HTML with layout…
Release Details and Core Specifications
Cohere officially released Parse on August 27, 2026. The announcement identifies the model by the ID parse-v5.0. It is described as a 2.3B-parameter multimodal model designed to turn complex documents into clean, structured Markdown ready for downstream AI workflows.
- Model ID: parse-v5.0
- Parameter count: 2.3B
- Context window: 8K tokens
- Approximate model size: 4.6GB
Extraction Capabilities and Output Formats
The source states that Parse extracts text in reading order, along with tables, lists, forms, images, captions, page boundaries, and visual element locations. The output includes Markdown/HTML content, HTML-formatted tables, bounding boxes, and image descriptions. The documentation notes that these outputs preserve both document structure and layout, which supports easier rendering and processing.
Evaluation Procedure and Source Limits
To evaluate the model, an engineer should test it against a diverse set of complex documents containing mixed layouts. Measure the accuracy of extracted tables and the precision of bounding boxes for visual elements. Compare the generated Markdown structure against manual annotations to validate reading order and page boundary detection. Validate that image descriptions align with the visual content provided in the input.
- Test with documents containing dense tables and multi-column layouts.
- Measure extraction fidelity for forms and lists.
- Compare output structure against ground-truth data.
- Validate bounding box coordinates against original image dimensions.
Primary sources
Frequently asked questions
What specific file formats does Parse process?
The source describes the model as turning complex documents into structured Markdown. It mentions extracting text, tables, forms, images, and captions. However, the provided facts do not list specific input file extensions such as PDF or DOCX.
Are the model weights available for local download?
The source records that open weights are not confirmed. It does not establish whether local installation is possible or if API access is mandatory. No pricing or availability statements are included in the verified facts.