typhoon-ai/typhoon-ocr1.5-2b

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Nov 10, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Typhoon-OCR-1.5-2B by typhoon-ai is a 2 billion parameter vision-language model built on Qwen3-VL, specifically designed for robust and fast Optical Character Recognition (OCR) of real-world documents in English and Thai. This model excels at understanding handwritten content, complex forms, and diverse layouts, providing structured outputs in Markdown, HTML, LaTeX, and figure formats. It features a compact architecture, faster inference without PDF metadata dependency, and a simplified single-prompt inference process, making it efficient for document intelligence tasks.

Loading preview...

Typhoon-OCR-1.5-2B: Advanced Vision-Language OCR for Thai and English Documents

Typhoon-OCR-1.5-2B is the latest iteration of typhoon-ai's open-source vision-language model, optimized for Optical Character Recognition (OCR) of real-world documents in both English and Thai. Built upon the Qwen3-VL 2B architecture, this model offers a significantly smaller footprint while enhancing multimodal capabilities and inference speed.

Key Enhancements & Capabilities

  • Compact and Efficient: Based on Qwen3-VL 2B, it's designed for efficient operation even on lightweight hardware, aided by quantization optimizations.
  • Faster Inference: Achieves high layout fidelity directly from image data, eliminating the need for PDF metadata and resulting in much faster processing for both PDFs and images.
  • Simplified Single-Prompt Inference: Replaces the previous two-prompt system with a single-prompt architecture, simplifying integration and ensuring consistent outputs across diverse document types.
  • Enhanced Understanding: Significantly improved at handling handwritten content, complex forms, and irregular layouts, interpreting elements with greater consistency and semantic accuracy.
  • Balanced Performance: Intelligently adapts its parsing pipeline for both text-rich and image-rich documents, from financial tables and academic papers to diagrams and handwritten notes.

Structured Output Format

Typhoon-OCR-1.5-2B produces machine-friendly structured outputs, ideal for downstream AI and document intelligence tasks:

  • Markdown: For general text.
  • HTML: For tables, including complex layouts.
  • Figure <figure>: For figures, charts, and diagrams, with descriptive text.
  • LaTeX: For mathematical equations.
  • Page number <page_number>: To preserve page numbering.

Intended Uses & Limitations

This is a task-specific model intended for use with its provided prompts. It does not include guardrails or VQA capabilities. Users should be aware of potential hallucinations inherent to LLMs and assess risks for their specific use cases. For local deployment, vLLM is recommended for GPU inference, and Ollama for on-device use.

Finetunes
1 models