typhoon-ai/typhoon-ocr-3b

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:May 27, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Typhoon-OCR-3B by typhoon-ai is a 3 billion parameter bilingual vision-language model, inspired by Qwen2.5-VL-Instruction, specifically designed for real-world document parsing in Thai and English. It excels at extracting structured and layout-heavy information from documents like financial reports, academic papers, receipts, and menus, outputting in Markdown or HTML formats. The model features multi-layered interpretation for figures, including observation, context analysis, text recognition, and artistic analysis, to provide comprehensive structured descriptions.

Loading preview...

Typhoon-OCR-3B: Bilingual Document Parsing

Typhoon-OCR-3B is a 3 billion parameter vision-language model developed by typhoon-ai, optimized for parsing real-world documents in both Thai and English. Inspired by models like olmOCR and based on Qwen2.5-VL-Instruction, it specializes in extracting information from diverse document types, from structured financial reports to layout-heavy receipts.

Key Capabilities

  • Bilingual Document Understanding: Excels in processing documents containing both Thai and English text, including complex layouts and mixed-language content.
  • Structured Document Parsing: Handles financial reports, academic papers, books, and government forms, outputting general text in Markdown and tables in HTML (supporting merged cells and complex layouts).
  • Layout-Heavy Document Processing: Efficiently extracts information from informal documents such as receipts, menus, tickets, and infographics, presenting them in Markdown with embedded tables.
  • Multi-layered Figure Interpretation: Provides detailed analysis of figures, charts, and diagrams through observation, context analysis, text recognition (Thai/English), and artistic/structural analysis, generating structured descriptions.
  • Performance: Outperforms GPT-4o and Gemini 2.5 Flash in Thai document understanding, particularly for complex layouts, as detailed in its Technical Report.

Good For

  • Developers needing robust OCR and document parsing for real-world Thai and English documents.
  • Applications requiring structured data extraction from various document formats, including those with complex visual elements.
  • Use cases where accurate interpretation of figures and tables within documents is critical.

Note: This model is task-specific and requires a specific prompt structure for optimal performance. It is not intended for general VQA or arbitrary prompting.

Finetunes
1 models