zfz04/FocusVTC

FocusVTC: Vision-Language Model for Document Understanding

FocusVTC is a 9 billion parameter vision-language model (VLM) based on the Qwen3.5 architecture, developed by zfz04. It is specifically engineered for advanced visual document understanding, particularly in scenarios involving long and multi-page documents. A key differentiator is its ability to perform evidence-focused tool use, allowing it to dynamically request higher-resolution crops of document regions when needed for more detailed analysis.

Key Capabilities

• Long-Context Visual Document Understanding: Optimized for processing and reasoning over extensive visual documents, including multi-page layouts.
• Iterative Region Zooming: Can identify potentially unreadable regions and utilize a zoom_region tool to obtain and process higher-resolution crops, enhancing fine-grained analysis.
• Visual Tool Use: Designed to integrate into an agent loop, enabling it to interact with external tools for visual retrieval and inspection.
• Fine-Grained Visual Evidence Localization: Excels at pinpointing specific visual evidence within documents.

Good For

• Long and Multi-page Document Question Answering: Ideal for extracting information and answering queries from complex, lengthy documents.
• Research on Multimodal Agents: Provides a strong foundation for developing agents that require sophisticated visual reasoning and tool interaction.
• Visual Retrieval and Inspection: Useful for applications needing detailed visual analysis and iterative refinement of visual observations.

Limitations

• Requires an external runtime for tool execution. The model itself does not automatically execute zoom_region calls.
• Performance can be influenced by factors like page resolution, prompt format, and generation settings.
• Outputs should always be independently verified, especially in high-stakes applications.