Granite Vision 3.1

What is granite-vision-3.1-2b-preview?

Granite Vision 3.1 is a sophisticated vision-language model specifically engineered for enterprise document understanding and analysis. Built on IBM's Granite LLM architecture, this model excels at processing visual documents, including tables, charts, infographics, and diagrams. With 3 billion parameters, it achieves state-of-the-art performance across various document understanding benchmarks while maintaining computational efficiency.

Implementation Details

The model architecture combines three key components: a SigLIP vision encoder, a two-layer MLP vision-language connector with GELU activation, and the granite-3.1-2b-instruct language model featuring 128k context length. The implementation builds upon LLaVA architecture with enhanced multi-layer encoder features and denser grid resolution for improved document interpretation.

• Trained on IBM's Blue Vela supercomputing cluster with H100 GPUs
• Leverages both public datasets and synthetic data for document understanding
• Implements advanced safety features through Granite Guardian integration

Core Capabilities

• Superior performance on ChartQA (0.86) and DocVQA (0.88) benchmarks
• Enhanced OCR capabilities with 0.75 score on OCRBench
• Strong general VQA performance (0.81 on VQAv2)
• Specialized in table and chart analysis
• Document content extraction and comprehension