IBM has introduced Granite 4.0 3B Vision, a compact vision-language model designed to parse complex enterprise documents, forms, and financial charts. Published under the open-source Apache 2.0 license on Hugging Face, the 3-billion-parameter model delivers high performance on document understanding while operating on modest compute footprints. Developers can deploy the model standalone or integrate it directly into existing document pipelines to extract structured data from visual media.
DeepStack Architecture and Modular Design
The model addresses key vision-language bottlenecks using IBM's DeepStack architecture. Instead of injecting visual features at a single entry point, DeepStack routes abstract visual features into early model layers for semantic context while injecting high-resolution spatial details into later layers. This design preserves both overall context and spatial alignment across dense document layouts.
Granite 4.0 3B Vision is packaged as a LoRA adapter built on top of the Granite 4.0 Micro dense language model. This modular setup allows enterprises to reuse a single deployment for both multimodal and text-only tasks, automatically falling back to the base language model when image processing is not required. Model training relied on ChartNet, a synthetic dataset of 1.7 million chart samples generated via code-guided pipelines across 24 chart types.
Key Performance Benchmarks
Across document understanding evaluations, IBM's compact model demonstrates strong performance against larger vision models:
- Chart Understanding: Granite 4.0 3B Vision achieves an 86.4% Chart2Summary score on ChartNet, outperforming all evaluated peer models.
- Table Extraction: Evaluated with the Tree Edit Distance Score (TEDS), it scored 92.1 on PubTables-v2 cropped tables, 79.3 on full-page documents, 64.0 on OmniDocBench, and 88.1 on TableVQA.
- Key-Value Pair Extraction: The model recorded an 85.5% exact match zero-shot accuracy score on the VAREX benchmark, which evaluates 1,777 U.S. government forms.
Implementation and Operational Caveats
For enterprise workloads, Granite 4.0 3B Vision supports standalone execution for single-image tasks like receipt parsing or integrated processing via the Docling framework for multi-page PDFs. When combined with Docling, the pipeline automatically detects, segments, and crops tables or figures before passing clean visual regions to the model for detailed extraction.
Despite its strengths, the model exhibits clear capacity trade-offs due to its small parameter count. On the Chart2CSV conversion metric, Granite 4.0 3B Vision scored 62.1%, trailing behind the larger Qwen3.5-9B model, which achieved 63.4%. Organizations prioritizing raw numerical conversion from dense charts may still require larger model alternatives.