Cohere Parse 5 Document Parsing Model Launch

Cohere officially released Parse 5, a dedicated 2.3 billion parameter vision-language model (VLM) for parsing unstructured enterprise documents like PDFs, PPTs, and JPEGs into clean, structured Markdown. It supports a 8192 context window and nine languages. The model performs end-to-end single-pass extraction, outputting text in correct reading order, tables as structured HTML, key-value pairs from forms, and image descriptions with bounding boxes, while preserving semantic formatting like strikethroughs and superscripts. It offers two output modes: full-page markdown and structured content blocks. Cohere claims a ParseBench score of 79.2, surpassing Mistral OCR 4 (74.5) and Azure Document Intelligence (69.3) on three of five dimensions, though LlamaParse Agentic leads on the full five-dimension benchmark. The evaluation excludes Layout and Charts dimensions, with chart data extraction planned for a future version.
Key insights
Parse 5 Aims to Solve the Bottleneck of Converting Unstructured Documents to LLM-Ready Data
The launch is framed around the idea that document parsing is a primary bottleneck preventing LLMs from effectively using enterprise data. Previously, organizations had to choose between high-quality, unscalable, and expensive parsing or scalable but low-quality, lossy parsing. Parse 5 is positioned as a solution that provides high-quality, scalable parsing at a low per-page cost, making it suitable for high-volume enterprise workloads. It is described as an end-to-end model with no separate OCR stage, which allows it to preserve complex document structures in a single forward pass.
Parse 5's Performance Claims Are Strong but Come with Specific Benchmarking Caveats
Cohere's claim of market-leading performance is based on a ParseBench score of 79.2, but this is measured on only three of the benchmark's five dimensions (Tables, Content Faithfulness, Semantic Formatting). The Layout and Charts dimensions were excluded—Layout because Parse 5 intentionally does not output bounding boxes for text blocks, and Charts because the model treats charts as images with descriptions rather than extracting their underlying data. When measured across all five dimensions, LlamaParse Agentic remains the leader with a score of 84.9. The benchmark's scoring rules were recently fixed to address a bug in bold/heading detection, and Cohere re-ran all competitors with the new rules.