IBM launches Granite 4.0 3B Vision for documents

IBM has released Granite 4.0 3B Vision, a compact AI model capable of interpreting tables, charts, and forms inside documents. It is designed to turn complex images and PDFs into structured data that other systems can use.

The model can extract tables as HTML, transform charts into CSV data or code, and locate key-value pairs such as "Invoice number: 45821" or "Due date: April 15." It can also describe images in natural language.

What it can do

Granite 4.0 3B Vision is designed for tasks where reading the text is not enough. You also need to understand the position of each element and the relationships between them.

• Tables: identifies rows, columns, and complex structures, even when a table appears on a page filled with other content.
• Charts: interprets lines, bars, and other formats to generate summaries, structured data, or executable code.
• Document fields: extracts information from invoices, receipts, forms, and documents with varied layouts.

For example, in a financial report, it can detect a chart, read its values, and convert them into a table that can then be analyzed automatically. In a form, it can identify which text corresponds to the name, address, or date, even when the layout changes from one document to another.

A small model with a modular architecture

Granite 4.0 3B Vision works as a LoRA adapter on top of Granite 4.0 Micro. A LoRA adapter adds specific capabilities to a base model without having to replace it entirely.

In practice, this makes it possible to use the same deployment for text and image tasks. When a request does not require vision, the system can rely on the base model. This separation reduces the complexity of integrating the model into business processes that combine documents, text, and images.

IBM can also combine it with Docling, a tool for analyzing documents. Docling detects and crops tables, charts, and figures inside multi-page PDFs, and Granite Vision then processes each element in greater detail.

The result is a workflow that can automate the reading of large document collections without sending entire pages to the model at every step.

How it improves chart interpretation

Charts are difficult for many vision models because they require combining three things: visual patterns, numerical values, and language. It is not enough to recognize that an image contains an upward-trending line. The system must know what each axis represents and, when possible, read the correct values.

To train this capability, IBM developed ChartNet, a dataset with 1.7 million charts across 24 types and six generation libraries. Each example combines five related elements:

• The code used to create the chart.
• The rendered image.
• The original data table.
• A natural-language summary.
• Questions and answers about its content.

According to IBM's published evaluations, Granite 4.0 3B Vision achieved 86.4% on Chart2Summary, the highest score among the models compared in that benchmark, which was evaluated using another language model. On Chart2CSV, it reached 62.1%, behind Qwen3.5-9B, which scored 63.4%.

The model also uses a variant of DeepStack Injection. This technique introduces abstract visual information into the model's early layers and high-resolution spatial details into later layers. This allows it to work simultaneously with the meaning of a page and the exact location of its elements.

Results on tables and forms

For table extraction, IBM reports the following TEDS scores, a metric that measures both the structure and the extracted content:

• 92.1 on cropped tables from PubTables-v2.
• 79.3 on full pages from PubTables-v2.
• 64.0 on tables from full pages in OmniDocBench.
• 88.1 on TableVQA-extract.

On the VAREX benchmark, which consists of 1,777 forms from the United States government, it achieved 85.5% accuracy without prior task-specific examples. In this test, an answer counts as correct only if the key-value pairs exactly match the reference data.

What this changes for you

The main value is not chatting with an image, but automating repetitive document work. A company could use it to process invoices, review financial reports, or make the visual information in academic documents searchable.

It can also power small, specialized tools, such as a form reader or chart analyzer, without requiring you to build an entire platform. For large volumes, integration with Docling makes it possible to work with multi-page PDFs and send the model only the relevant sections.

Granite 4.0 3B Vision is already available on Hugging Face under the Apache 2.0 license. The important next step will be seeing how it performs on real-world documents, low-quality scans, and layouts that do not appear in its benchmarks. The direction is clear: document models are moving beyond reading words and starting to turn a page's visual structure into usable data.