typhoon-ai/typhoon-ocr1.5-3b-qat

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Nov 10, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Typhoon-OCR-1.5-3B-QAT by typhoon-ai is a 3 billion parameter vision-language model with a 32768 token context length, specifically designed for robust and efficient on-device Optical Character Recognition (OCR). This model is a quantization-aware trained (QAT) version of Typhoon OCR v1.5, built upon Qwen2.5-VL-3B, ensuring strong accuracy even under aggressive low-bit quantization. It excels at extracting text from diverse documents in both English and Thai, including handwriting, forms, and complex layouts, making it ideal for lightweight, resource-constrained environments.

Loading preview...

Typhoon-OCR-1.5-3B-QAT Overview

Typhoon-OCR-1.5-3B-QAT is a 3 billion parameter vision-language model developed by typhoon-ai, specifically optimized for Optical Character Recognition (OCR) in English and Thai. This model is a Quantization-Aware Trained (QAT) version of Typhoon OCR v1.5, built on the Qwen2.5-VL-3B architecture. Its primary differentiator is its ability to maintain high OCR accuracy even when deployed with aggressive low-bit quantization (e.g., 4-bit), making it highly efficient for on-device inference in resource-constrained environments.

Key Capabilities

  • Quantization-Aware Training (QAT): Engineered to preserve strong OCR accuracy and stability when converted to low-bit formats, significantly reducing performance degradation.
  • Optimized for On-Device Inference: Delivers faster and more consistent performance on hardware with limited resources.
  • Enhanced Document Understanding: Retains and improves upon v1.5's capabilities in handling challenging document types, including handwritten notes, complex forms, irregular layouts, and structured documents.
  • Versatile Document Support: Effectively processes text-rich and image-rich documents such as tables, diagrams, annotated pages, charts, receipts, and dense reports.
  • Thai + English Multilingual OCR: Provides reliable text extraction from bilingual real-world documents.

Good for

  • Edge Device Deployment: Ideal for applications requiring efficient OCR on mobile devices, embedded systems, or other low-resource hardware.
  • High-Volume OCR Workloads: Suitable for scenarios where maintaining accuracy under aggressive quantization is critical for throughput and cost efficiency.
  • Processing Complex Documents: Excellent for extracting information from diverse and challenging document types, including those with handwriting or intricate layouts.
  • Bilingual Thai and English OCR: Specifically designed for use cases involving documents containing both Thai and English text.

Finetunes
1 models