AI 軟體／文件處理
Cohere Parse 5 Uses a 2.3B Vision Model to Process Enterprise Documents, but Its Benchmark Score Excludes Charts and Layout Localization
Parse 5 converts PDFs, presentations, and images into Markdown while preserving tables and reading order, with API pricing of $1.50 per 1,000 pages. Its officially reported ParseBench score of 79.2 covers only three of the benchmark’s five evaluation dimensions and should not be treated as a complete document-understanding score.
Cohere has released the Parse 5 document parsing model, identified as `parse-v5.0`. It is a proprietary vision-language model with approximately 2.3 billion parameters, a size of 4.6GB, and an 8,192-token context window. It can extract text, reading order, tables, forms, lists, images, and captions from PDFs, PPT files, and images, then output Markdown or ordered content blocks. Tables and images can include bounding boxes. This allows it to sit directly at the entry point of RAG, document indexing, or agent workflows, reducing the engineering required to integrate OCR, layout analysis, and format reconstruction separately.
The service is now available through the Cohere API, Model Vault, Microsoft Foundry, and AWS SageMaker. The Cohere API is priced at $1.50 per 1,000 pages, and the company says eight H100 GPUs can process 36 pages per second. However, this throughput was measured in a vendor-controlled environment; actual latency will still depend on page resolution, the proportion of pages routed through the vision path, batch size, and retries. The model’s weights are not publicly available, and it does not return confidence scores for extracted content. Regulated workflows will therefore still require separate sampling, field-validation, and human-review mechanisms.
The evaluation methodology deserves the closest scrutiny. Cohere reran ParseBench under updated rules and achieved an average score of 79.2 across three categories: tables, content fidelity, and semantic formatting. That score trails GPT-5.5, Opus 4.8, and Gemini 3.5 Flash, but slightly exceeds LlamaParse Cost Effective. The original ParseBench, however, includes approximately 2,000 pages, more than 169,000 rules, and five dimensions. Cohere excluded chart data extraction and full visual localization from its average because Parse currently only describes charts and does not output a bounding box for every text segment. Engineering teams should therefore build regression sets using their own Chinese-language scans, multi-page tables, seals, and charts rather than comparing a single aggregate score.