QCRI/Fanar-2-Oryx-IVU

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Mar 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

QCRI/Fanar-2-Oryx-IVU is a 7 billion parameter Arabic-first vision-language model developed by Qatar Computing Research Institute (QCRI) at Hamad Bin Khalifa University (HBKU). Built on Qwen2.5-VL-Instruct, it specializes in culturally-aware image and video understanding, Arabic calligraphy recognition, and strong Arabic language support. The model is fine-tuned on 62 million bilingual examples, achieving 70% user satisfaction and significantly reduced code-switching in Arabic responses, making it ideal for applications requiring deep cultural and linguistic understanding of Arabic visual content.

Loading preview...

Fanar-2-Oryx-IVU: Arabic-First Vision-Language Model

Fanar-2-Oryx-IVU is a 7 billion parameter vision-language model (VLM) developed by the Qatar Computing Research Institute (QCRI) at Hamad Bin Khalifa University (HBKU). It is an Arabic-first model designed for culturally-aware image and video understanding, building upon the Qwen2.5-VL-Instruct base model. This model is a key component of the Fanar 2.0 release, a comprehensive Arabic-centric multimodal generative AI platform.

Key Capabilities

  • Culturally-Aware Understanding: Specializes in interpreting images and videos with strong Arabic language support and cultural context, trained on 62 million balanced Arabic-English multimodal examples.
  • Arabic Calligraphy Recognition: Identifies content and script styles (e.g., Thuluth, Naskh) across 5 major Arabic scripts, trained on 20,000 calligraphy images.
  • Reduced Code-Switching: Achieves a 45% reduction in Arabic-English code-switching and a 50% reduction in Arabic-Chinese mixing compared to its base model, leading to more coherent Arabic responses.
  • High User Satisfaction: Demonstrates 70% user satisfaction in generative evaluations, outperforming other 7B-class models in quality.
  • Point-Based Grounding: Utilizes (x,y) coordinates for enhanced spatial reasoning and object localization.

Good For

  • Cultural Heritage & Education: Documenting and preserving cultural heritage, and educational applications for Arabic culture and history.
  • Arabic Content Moderation: Analyzing and moderating Arabic social media content.
  • E-commerce & Tourism: Generating Arabic product descriptions and supporting multilingual tourism applications.
  • Calligraphy & Document Analysis: Analyzing historical texts and various Arabic script styles.
  • Research: Advancing Arabic vision-language understanding and culturally-aware AI systems.

Finetunes
1 models