Document Monster 3B

Granite 4.0 3B Vision is a small multimodal VLM tailored for enterprise document understanding. It focuses on reliably extracting information from complex documents, forms, and structured visual materials.

The key capabilities are as follows.

• Table Extraction: extracts multi-row, multi-column table structures from document images
• Chart Understanding: converts charts and figures into structured data, summaries, and executable code
• Semantic KVP Extraction: detects meaning-based key-value pairs across diverse layouts

The model is provided as a LoRA adapter on top of Granite 4.0 Micro. This allows a natural fallback to the text-only base model when vision isn't needed, making it easy to fit both multimodal and text workloads into the same deployment flow.

The performance improvements center on three elements.

• ChartNet: a 1.7 million-scale synthetic multimodal dataset built for chart interpretation
• DeepStack Injection: an architectural variant that injects low-level visual features and high-level semantics into different layers
• Modular design: separates vision/language to suit enterprise deployment

On benchmarks, it delivered strong results in chart and table extraction.

• Chart2Summary 86.4%, the top score among evaluated models
• Chart2CSV 62.1%, just below the larger Qwen3.5-9B (63.4%)
• In table extraction, it leads with PubTables-v2 cropped 92.1, full-page 79.3, OmniDocBench 64.0, and TableVQA 88.1
• On the VAREX KVP benchmark, it recorded zero-shot 85.5% EM

There are two usage patterns.

• Stand-alone image understanding used directly on a single image
• A document pipeline combined with Docling that processes entire PDFs, cropping tables and figures to pass to Granite Vision

Suggested use cases include invoice/form processing, financial report analysis, and research document intelligence. The license is Apache 2.0.