Hugging Face has announced Granite 4.0 3B Vision, a compact vision‑language model released as a LoRA adapter on top of its dense language model Granite 4.0 Micro. The design keeps vision and language modular, allowing the same deployment to serve both multimodal and text‑only workloads with automatic fallback to the base model when vision is not required. Granite 4.0 3B Vision is purpose‑built for reliable information extraction from complex documents, forms, and structured visuals, and can be used either as a stand‑alone visual extraction engine or integrated with Docling to support end‑to‑end document understanding pipelines. The model’s performance is driven by three key investments: a purpose‑built chart understanding dataset called ChartNet, a novel variant of the DeepStack architecture that injects abstract visual features into earlier layers and high‑resolution spatial features into later layers, and a modular design that keeps the model practical for enterprise deployment. ChartNet is a million‑scale multimodal dataset that includes 1.7 million chart samples across 24 chart types and six plotting libraries, each sample comprising plotting code, rendered image, data table, natural language summary, and QA pairs. Granite 4.0 3B Vision achieves the highest Chart2Summary score of 86.4 % on the human‑verified ChartNet benchmark and ranks second on Chart2CSV with 62.1 %. In table extraction, the model leads on PubTables‑V2 (92.1 % cropped, 79.3 % full‑page), OmniDocBench (64.0 %) and TableVQA (88.1 %) using the TEDS metric. On the VAREX benchmark of U.S. government forms, Granite 4.0 3B Vision attains 85.5 % exact‑match accuracy zero‑shot. The model is available on Hugging Face under the Apache 2.0 license, with full technical details, training methodology, and benchmark results documented in the model card.
Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents
Hugging Face has released Granite 4.0 3B Vision, a compact vision‑language model built as a LoRA adapter on top of Granite 4.0 Micro. The model is designed for enterprise document understanding, offering reliable extraction from complex documents, forms, and structured visuals. It supports tasks such as detailed image descriptions and can be used standalone or integrated with Docling for full document pipelines. Granite 4.0 3B Vision achieves strong performance on chart interpretation, table extraction, and key‑value pair extraction, outperforming many larger models on benchmark datasets.
Why it matters
The model demonstrates that compact, modular VLMs can match or exceed larger counterparts on enterprise‑relevant tasks, enabling more efficient deployment in business settings.