Granite Vision 3.3 2B

Granite-vision-3.3-2b is a vision-language model designed for visual document understanding, enabling automated content extraction from tables, charts, infographics, plots, diagrams, and general images. It is a compact model with approximately 2.97 billion total parameters, built by fine-tuning a 2-billion-parameter Granite large language model with a SigLIP2 vision encoder and a two-layer MLP connector.

Key Strengths

The model excels at document understanding tasks, achieving strong benchmark results compared to prior Granite Vision versions:

Benchmark	Granite-vision-3.3-2b	Granite-vision-3.2-2b
DocVQA	0.91	0.89
ChartQA	0.87	0.87
TextVQA	0.80	0.78
InfoVQA	0.68	0.64
OCRBench	0.79	0.77

Safety alignment has been improved across all measured dimensions. On the RTVLM benchmark, the model scores 8.0 (Politics), 8.1 (Racial), 7.5 (Jailbreak), and 8.0 (Mislead). On VLGuard, it scores 8.4 for unsafe images and 9.3 for safe images with unsafe instructions.

Experimental Capabilities

The model introduces three experimental features: image segmentation, doctags generation for parsing document images into structured text, and multi-page support for question answering across up to 8 consecutive document pages.

Architecture and Training

The model uses a SigLIP2 vision encoder, a two-layer MLP vision-language connector, and the granite-3.1-2b-instruct language model with 128k context length. Training data includes publicly available datasets, internally created synthetic data for document understanding, and high-quality sources such as Mammoth-12M and Bigdocs. The model is released under the Apache 2.0 license.