Granite Vision 3.2

Granite Vision 3.2 2B is IBM's lightweight vision-language model, released in February 2025 as part of Granite 3.2. Built with IBM's Docling toolkit processing 85 million PDFs plus 26 million synthetic QA pairs, it targets visual document understanding: extracting data from tables, charts, infographics, scanned pages, and diagrams.

Despite only 2 billion parameters, it matches or outperforms much larger open models like Llama 3.2 11B and Pixtral 12B on enterprise benchmarks such as DocVQA, ChartQA, AI2D, and OCRBench. Released under Apache 2.0 with weights, GGUF quantizations, and training transparency consistent with IBM's open Granite strategy.