Granite Vision 4.1 4B
Granite Vision 4.1 4B is an image-and-text model for document understanding. Its approximately 4B parameters include a 3.4B language component and about 0.6B in the visual encoder and projectors.
Model design and capabilities
Layer Deepstack inserts features from four visual depths into the language model. The deepest visual features enter earlier language layers. The model targets chart-to-CSV, chart descriptions, table JSON and key-value extraction. The card specifies English instructions.Granite Vision 4.1 4B model card
Practical use
Practical guidance. Use this model when document structure matters to the answer. Review extracted cell values and labels against the source image, especially in dense tables. Compare a structured output with a visual description; fluent prose can conceal a missed row or an incorrect axis.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | ibm-granite/granite-vision-4.1-4b |
| License | Apache 2.0Granite Vision 4.1 4B model card |
| Language backbone | Fine-tuned Granite 4.1 3B.Granite Vision 4.1 4B model card |
| Vision encoder | SigLIP2 SO400M, with 384-pixel tiles.Granite Vision 4.1 4B model card |
| Hosted access | watsonx: deploy on demand.IBM watsonx.ai foundation model library |
Related models
- Granite Docling 258M
IBM ยท Granite Docling
A document conversion model that turns page images into structured DocTags text.
Launch date not confirmedDownloadable