今天,我们激动地宣布推出 Granite 4.0 3B Vision,这是一款专为企业文档理解而设计的紧凑型视觉语言模型(VLM)。它专为从复杂文档、表单和结构化视觉内容中可靠提取信息而构建。Granite 4.0 3B Vision 在以下能力上表现出色:
- 表格提取:从文档图像中精确解析复杂的表格结构(例如,多行、多列等)
- 图表理解:将图表和图形转换为结构化的机器可读格式、摘要或可执行代码
- 语义键值对(KVP)提取:在多种文档布局中识别并定位具有语义意义的键值字段对
该模型以 LoRA 适配器的形式部署在 Granite 4.0 Micro(我们的密集语言模型)之上,保持视觉和语言模块的独立性,以便支持纯文本回退并无缝集成到混合流水线中。它继续支持视觉语言任务,例如根据图像生成详细的自然语言描述(例如,“详细描述这张图片”)。该模型可以独立使用,也可以与 Docling 协同使用,通过深度视觉理解能力增强文档处理流水线。
Granite 4.0 3B Vision 的构建方式
Granite 4.0 3B Vision 的性能源于三项关键投入:一个通过新颖的代码引导数据增强方法构建的专用图表理解数据集、一种能够注入高细节视觉特征的新型 DeepStack 架构变体,以及一种使模型在企业部署中保持实用性的模块化设计。
ChartNet:教会模型真正理解图表
图表对视觉语言模型(VLM)构成了挑战,因为理解图表需要同时对视觉模式、数值数据和自然语言进行联合推理——这种组合大多数 VLM 都难以妥善处理,尤其是在空间精度至关重要的情况下,例如从折线图中读取精确数值。为了弥补这一差距,我们开发了 ChartNet:一个百万级的多模态数据集,专为图表解读和推理而构建,详细信息将在我们即将发表的 CVPR 2026 论文中阐述。
ChartNet 采用代码引导的合成流程,生成了 170 万个多样化的图表样本,涵盖 24 种图表类型和 6 个绘图库(见图 1)。其独特之处在于,每个样本由五个对齐的组件组成——绘图代码、渲染图像、数据表格、自然语言摘要和问答对——这为模型提供了对图表含义的深度跨模态理解,而不仅仅是图表的外观。该数据集还包含了人工标注和真实世界的子集,并经过视觉保真度、语义准确性和多样性的筛选。
其成果是一个训练资源,它使视觉语言模型从仅仅描述图表,转变为真正理解图表所编码的结构化信息——在不同模型规模、架构和任务上均取得了一致的性能提升。
图 1:ChartNet 的合成数据生成流程。
DeepStack:更智能的视觉特征注入
大多数视觉语言模型在单一节点将视觉信息注入其语言模型,这迫使模型同时处理高层语义和细粒度空间细节。Granite 4.0 3B Vision 采用了一种不同的方法,即 DeepStack 注入:抽象的视觉特征被路由到较早的层以进行语义理解,而高分辨率空间特征则被馈送到较晚的层以保留细节。其结果是,模型既能理解文档中的内容,也能理解其位置——这对于表格提取、图表理解和键值对解析等布局与内容同等重要的任务至关重要。有关完整的技术细节,请参阅模型卡片中的模型架构部分。
模块化:一个模型,两种模式
Granite 4.0 3B Vision 被打包为 Granite 4.0 Micro 之上的一个 LoRA 适配器,而非一个独立的模型。在实践中,这意味着同一个部署可以同时处理多模态和纯文本工作负载,并在不需要视觉功能时自动回退到基础模型。这使得企业集成变得简单直接,同时不牺牲性能。
性能表现如何
图表:在基于 LLM 作为评判者的人工验证 ChartNet 基准测试上,Granite 4.0 3B Vision 在所有评估模型中(包括参数量显著更大的模型)取得了最高的 Chart2Summary 分数(86.4%)[见图 2]。其 Chart2CSV 分数(62.1%)也排名第二,仅次于参数量是其两倍以上的 Qwen3.5-9B(63.4%)。
图 2:Granite 4.0 3B Vision 在 chart2csv 和 chart2summary 任务上的性能,与同类视觉语言模型在 LLM 作为评判者的方法下进行对比。
表格:我们在两种设置下评估表格提取能力:裁剪表格(孤立区域)和整页文档(表格嵌入在复杂布局中)[见图 3]。基准测试套件包括 TableVQA-extract(裁剪表格图像)、OmniDocBench-tables(整页文档)以及 PubTables-v2(同时包含裁剪和整页设置)。模型的任务是以 HTML 格式提取表格,并使用 TEDS 指标进行评分,该指标同时衡量结构准确性和内容准确性。Granite 4.0 3B Vision 在所有评估模型中表现最强,在 PubTablesV2 的裁剪(92.1)和整页(79.3)设置、OmniDocBench(64.0)以及 TableVQA(88.1)分数上均领先。
图 3:Granite 4.0 3B Vision 在裁剪和整页基准测试(TableVQA-extract、PubTables-v2、OmniDocBench-tables)上的表格提取性能,以 TEDS 指标衡量。
语义 KVP:VAREX 是一个专门为区分小型提取模型而设计的基准测试,包含 1,777 份美国政府表格,涵盖从简单的平面布局到复杂的嵌套和表格结构。模型使用精确匹配(EM)进行评估,这是一个严格的指标,要求模型提取的键值对与真实标注完全一致。Granite 4.0 3B Vision 在零样本情况下达到了 85.5% 的 EM 准确率。
如何使用
Granite 4.0 3B Vision 既可以作为独立的视觉信息提取引擎运行,也可以作为 Docling 全自动文档处理流程的一部分。该模型旨在支持跨多种文档类型和视觉格式的可扩展、高精度提取。
1. 独立图像理解 Granite 4.0 3B Vision 可直接在单张图像上运行,这一特性使其适用于那些已有工作流程、只需在不修改上游系统的情况下进行针对性视觉提取的应用场景。该功能可轻松集成到现有自动化工作流中,适合轻量级、任务特定的工具(例如表单解析器、图表分析器等)。
2. 集成文档理解流水线(配合 Docling) Granite 4.0 3B Vision 还可与 Docling 无缝集成,支持完整的端到端文档理解。该模式可提供:
- 多页 PDF 的大规模处理
- 利用 Docling 自动检测、分割并裁剪图表、表格及其他视觉元素,并将干净的裁剪结果重定向至 Granite Vision 模型进行细粒度提取
- 高效工作流,整体计算成本更低,吞吐量更快
- 更高的准确率、更可靠的提取效果,以及在大规模文档集合上显著提升的效率
示例用例
- 表单处理:利用键值对(KVP)能力从发票、表单和收据中提取结构化字段,或使用图像转文本(image2text)功能生成图像的自然语言描述(例如“详细描述这张图像”)。
- 财务报告分析:使用 Docling 解析报告、检测图表并裁剪视觉元素。利用 Granite Vision 的 chart2csv、chart2code 能力处理图表,并使用 tables_json 能力处理表格,将其转换为结构化、机器可读的数据,从而获得可操作的洞察。
- 研究文档智能:利用 Docling 处理密集学术 PDF 中的 OCR 和版面解析,并将提取出的图表传递给 chart2summary、将表格裁剪结果传递给 tables_html,使视觉内容与自由文本在同一流水线中均可被发现。
立即体验
Granite 4.0 3B Vision 现已通过 HuggingFace 发布,采用 Apache 2.0 许可证。完整的技术细节、训练方法和基准测试结果均可在模型卡片中查看。我们期待您用它构建的应用——欢迎在社区标签页中分享您的反馈。
Today we're excited to announce Granite 4.0 3B Vision, a compact vision-language model (VLM) designed for enterprise document understanding. It’s purpose-built for reliable information extraction from complex documents, forms, and structured visuals. Granite 4.0 3B Vision excels on the following capabilities:
- Table Extraction: Accurately parsing complex table structures (e.g., multi-row, multi-column, etc.) from document images
- Chart Understanding: Converting charts and figures into structured machine-readable formats, summaries, or executable code
- Semantic Key-Value Pair (KVP) Extraction: Identifying and grounding semantically meaningful key-value field pairs across diverse document layouts
The model ships as a LoRA adapter on top of Granite 4.0 Micro, our dense language model, keeping vision and language modular for text-only fallbacks and seamless integration into mixed pipelines. It continues to support vision-language tasks such as producing detailed natural-language descriptions from images (e.g., “Describe this image in detail”). The model can be used standalone or in tandem with Docling to enhance document processing pipelines with deep visual understanding capabilities.
How Granite 4.0 3B Vision Was Built
Granite 4.0 3B Vision’s performance is the result of three key investments: A purpose-built chart understanding dataset constructed via a novel code-guided data augmentation approach, a novel variant of the DeepStack architecture that enables high-detail visual feature injection, and a modular design that keeps the model practical for enterprise deployment.
ChartNet: Teaching Models to Truly Understand Charts
Charts present a challenge for vision-language models (VLMs) because understanding them requires jointly reasoning over visual patterns, numerical data, and natural language, a combination most VLMs cannot handle well, especially when spatial precision matters—such as reading exact values off a line chart. To close this gap, we’ve developed ChartNet: a million-scale multimodal dataset purpose-built for chart interpretation and reasoning, described in detail in our upcoming CVPR 2026 paper.
ChartNet uses a code-guided synthesis pipeline to generate 1.7 million diverse chart samples spanning 24 chart types and 6 plotting libraries [see Figure 1]. What makes it so distinctive is that each sample consists of five aligned components—plotting code, rendered image, data table, natural language summary, and QA pairs—providing models a deeply cross-modal view of what a chart means, not just what it looks like. The dataset also includes human-annotated and real-world subsets, filtered for visual fidelity, semantic accuracy, and diversity.
The result is a training resource that moves VLMs from merely describing charts to genuinely understanding the structured information they encode—with consistent gains across model sizes, architectures, and tasks.
Figure 1: ChartNet’s synthetic data generation pipeline.
DeepStack: Smarter Visual Feature Injection
Most VLMs inject visual information into their language model at a single point, which forces the model to handle both high-level semantics and fine-grained spatial detail simultaneously. Granite 4.0 3B Vision takes a different approach with DeepStack Injection: abstract visual features are routed into earlier layers for semantic understanding, while high-resolution spatial features are fed into later layers to preserve detail. The result is a model that understands both what is in a document and where—which is critical for tasks like table extraction, chart understanding, and KVP parsing where layout matters as much as content. For a full technical breakdown, see the Model Architecture section of the model card.
Modularity: One Model, Two Modes
Granite 4.0 3B Vision is packaged as a LoRA adapter on top of Granite 4.0 Micro, rather than as a standalone model. In practice, this means the same deployment can serve both multimodal and text-only workloads, automatically falling back to the base model when vision isn’t required. This keeps enterprise integration straightforward without sacrificing performance.
How It Performs
Charts: Evaluated on the human-verified ChartNet benchmark using LLM-as-a-judge, Granite 4.0 3B Vision achieves the highest Chart2Summary (86.4%) score among all evaluated models, including significantly larger ones [see Figure 2]. It also ranks second on Chart2CSV (62.1%), behind only Qwen3.5-9B (63.4%), a model more than double its size.
Figure 2: Granite 4.0 3B Vision performance on chart2csv and chart2summary, compared against peer vision-language models using LLM-as-a-judge.
Tables: We evaluate table extraction in two settings: cropped tables (isolated regions) and full-page documents (tables embedded in complex layouts) [see Figure 3]. The benchmark suite includes TableVQA-extract (cropped table images), OmniDocBench-tables (full-page documents), and PubTables-v2 (both cropped and full-page settings). Models are tasked with extracting tables in HTML format and scored using TEDS, a metric that captures both structural and content accuracy. Granite 4.0 3B Vision achieves the strongest performance across benchmarks, leading on PubTablesV2 on both cropped (92.1) and full-page (79.3), OmniDocBench (64.0), and TableVQA (88.1) scores among all evaluated models.
Figure 3: Granite 4.0 3B Vision’s table extraction performance across cropped and full-page benchmarks (TableVQA-extract, PubTables-v2, OmniDocBench-tables), measured by TEDS.
Semantic KVP: VAREX is a benchmark specifically designed to discriminate between small extraction models, comprising 1,777 U.S. government forms spanning simple flat layouts to complex nested and tabular structures. Models are evaluated using exact match (EM), a strict metric that requires the model’s extracted key-value pairs to match the ground truth. Granite 4.0 3B Vision achieves 85.5% EM accuracy zero-shot.
How to Use It
Granite 4.0 3B Vision can operate either as a stand‑alone visual information extraction engine or as part of a fully automated document‑processing pipeline with Docling. The model is designed to support scalable, accurate extraction across diverse document types and visual formats.
1. Stand‑Alone Image Understanding Granite 4.0 3B Vision can run directly on individual images, making this option useful for applications with existing workflows that need targeted visual extraction without modifying upstream systems. This offers easy integration into existing automation workflows and is suitable for lightweight, task‑specific tools (e.g., form parsers, chart analyzers, etc.).
2. Integrated Document Understanding Pipeline With Docling Granite 4.0 3B Vision can also be integrated seamlessly with Docling to support complete end‑to‑end document understanding. This mode can offer:
- Large‑scale processing of multi‑page PDFs
- Automated detection, segmentation, and cropping of figures, tables, and other visual elements with Docling and redirection of clean crops to Granite Vision model for fine-grained extraction
- Efficient workflow with lower overall computational costs and faster throughput
- Higher accuracy, more reliable extraction, and significantly improved efficiency across large document collections
Example Use Cases
- Form Processing: Extract structured fields from invoices, forms, and receipts using KVP capabilities or generate natural‑language descriptions of figures using image2text feature (e.g., “Describe this image in detail”).
- Financial Report Analysis: Use Docling to parse reports, detect figures, and crop visual elements. Process charts using Granite Vision’s chart2csv, chart2code, and tables using tables_json capabilities to convert them into structured, machine‑readable data enabling actionable insights.
- Research Document Intelligence: Utilize Docling to handle OCR and layout parsing across dense academic PDFs, and pass extracted figures to chart2summary and table crops to tables_html to make visual content discoverable alongside free-form text in a single pipeline.
Try It Today
Granite 4.0 3B Vision is available now on HuggingFace, released under the Apache 2.0 license. Full technical details, training methodology, and benchmark results are available in the model card. We’d love to hear what you build with it—share your feedback in the community tab.