Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents

IBM has released Granite 4.0 3B Vision, a compact 3-billion parameter vision-language model (VLM) purpose-built for enterprise document understanding. The model is designed to reliably extract information from complex documents, forms, and structured visuals, with specific capabilities in table extraction, chart understanding, and semantic key-value pair parsing. A key innovation is its modular design: it ships as a LoRA adapter on top of the text-only Granite 4.0 Micro model, allowing the same deployment to handle both multimodal and text-only workloads seamlessly.

Performance is driven by three core technical investments. First is ChartNet, a novel million-scale multimodal dataset for chart interpretation, containing 1.7 million diverse synthetic chart samples with aligned components like plotting code and data tables. Second is a novel variant of the DeepStack architecture, which injects abstract visual features into earlier model layers for semantic understanding and high-resolution spatial features into later layers to preserve critical layout details. This approach enables the model to understand both the 'what' and the 'where' in a document, which is essential for tasks like accurately reading values from a line chart or parsing a multi-column table.