IBM's 3B Granite Vision model beats much larger rivals at reading charts

IBM just dropped a small but surprisingly sharp vision-language model called Granite 4.0 3B Vision, and it punches way above its weight class on enterprise document tasks. The model is designed for the unglamorous but critical work of parsing complex tables, understanding charts, and extracting key-value pairs from messy document layouts — exactly the kind of thing that makes or breaks an enterprise AI workflow.

The secret sauce is a three-part recipe. First, the team built ChartNet, a 1.7 million-sample dataset that doesn’t just show the model what a chart looks like. Each sample bundles together plotting code, the rendered image, the underlying data table, a summary, and Q&A pairs. That cross-modal view teaches the model to genuinely understand structured information, not just describe it, as outlined in their upcoming CVPR 2026 paper. Second, they used a technique called DeepStack Injection that routes abstract visual features into earlier layers of the language model while pushing high-resolution spatial details into later ones. That split keeps the model aware of both what’s in a document and where it sits on the page.

The third piece is pure practicality. The model ships as a LoRA adapter on top of IBM’s Granite 4.0 Micro, so a single deployment can toggle between multimodal and text-only modes. No need to run a separate vision stack when you’re just doing text.

The numbers back up the architecture. On a human-verified chart benchmark, Granite 4.0 3B Vision scored 86.4% on Chart2Summary, beating every other model evaluated, including ones with far more parameters. It landed second on Chart2CSV at 62.1%, trailing only Qwen3.5-9B — a model more than double its size — by just 1.3 percentage points. On table extraction, it led across multiple benchmarks, hitting 92.1 on PubTablesV2 cropped tables. That a 3-billion-parameter model is competitive with 9-billion-parameter alternatives while using a modular, enterprise-friendly architecture suggests the real action in document AI isn’t about scaling up — it’s about training smarter.