Ovis2-2B
| Property | Value |
|---|---|
| Model Size | 2 Billion parameters |
| Architecture | aimv2-large-patch14-448 (Vision) + Qwen2.5-1.5B-Instruct (Language) |
| License | Apache License 2.0 |
| Paper | arXiv:2405.20797 |
What is Ovis2-2B?
Ovis2-2B is a state-of-the-art multimodal large language model that excels in combining visual and textual understanding. As part of the Ovis2 series, it represents a significant advancement in achieving structural alignment between visual and textual embeddings, delivering impressive performance despite its relatively compact size.
Implementation Details
The model combines an aimv2-large-patch14-448 vision transformer with Qwen2.5-1.5B-Instruct language model, optimized for efficient processing of both visual and textual inputs. It supports various input formats including single images, multiple images, videos, and text-only queries, with a maximum context length of 32,768 tokens.
- Advanced visual-text alignment architecture
- Optimized training methodology for small-scale model efficiency
- Comprehensive support for multiple input modalities
- Enhanced multilingual OCR capabilities
Core Capabilities
- Strong performance in benchmark tests (76.9% on MMBench-V1.1test)
- Superior OCR performance (87.3% on OCRBench)
- Robust video understanding capabilities
- Advanced reasoning through Chain-of-Thought prompting
- Multi-image and video frame processing
- Structured data extraction from tables and charts
Frequently Asked Questions
Q: What makes this model unique?
Ovis2-2B stands out for its ability to achieve high performance metrics comparable to larger models while maintaining a relatively small parameter count. Its structural embedding alignment approach enables superior visual-text understanding and reasoning capabilities.
Q: What are the recommended use cases?
The model excels in various applications including visual question answering, image and video description, OCR tasks, multilingual document processing, and complex reasoning tasks requiring visual context understanding.
