Sold by: IBM Data and AI
Deployed on AWS
IBM Granite Vision 3.3 2B is a compact multimodal model for accurate document image understanding, segmentation, and visual question answering.
Overview
IBM Granite Vision 3.3 2B is a lightweight, open-source multimodal model designed for enterprise-grade document image understanding. It excels at tasks like visual question answering, OCR, chart and table parsing, image segmentation, and structured document analysis across scanned documents, diagrams, and infographics. Built on the Granite 3.3 decoder-only LLM and enhanced with a SIGLIP2 vision encoder, it supports up to 8-page documents and generates structured outputs like doctags. Granite Vision is fine-tuned on a diverse mix of public and synthetic datasets to ensure high accuracy and robustness, especially on noisy, low-quality enterprise documents. It ranks among the top sub-7B models on OCRBench and delivers competitive results across DocVQA, TextVQA, and ChartQA. Released under Apache 2.0, it includes examples for MM-RAG, segmentation, and Hugging Face integration, and aligns with IBM's commitment to transparent, safe, and enterprise-ready AI.
Highlights
- Granite Vision 3.3 2B excels at understanding complex visual documents, including scanned pages, charts, diagrams, and tables. It supports visual question answering, OCR, and structured extraction through doctags. With strong performance on benchmarks like DocVQA, OCRBench, and TextVQA, it delivers accurate results even on low-quality or noisy enterprise documents, enabling reliable automation for document-heavy workflows.
- Granite Vision 3.3 2B uses an optimized architecture combining the SIGLIP2 vision encoder and Granite 3.3 language decoder to deliver high performance with a compact 2B parameter footprint. It supports up to 8-page documents and introduces capabilities like image segmentation and doctag generation. Despite its small size, it ranks among the top open-source models under 7B parameters on OCRBench, offering strong accuracy and efficiency for real-world enterprise use cases.
- Designed for diverse enterprise applications, Granite Vision 3.3 2B enables accurate visual understanding in workflows such as intelligent document processing, claims analysis, and compliance review. Its ability to handle complex layouts, extract structured data, and answer questions across multi-page documents makes it ideal for automating manual review tasks in industries like finance, insurance, legal, and healthcare.
Details
Sold by
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
This product is available free of charge. Free subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimensions summary
This model is free to use, so you pay only for the compute instance you run it on. Pricing is charged per host hour (HostHrs) and varies by instance type. Two batch options run on ml.g5 instances for offline processing. Eight real-time options run on ml.g6e, ml.p4d, and ml.p5 instances for live inference. Within each mode, larger instance sizes carry more compute capacity. You choose the instance that fits your workload and processing mode. Costs scale with how long each instance runs and which instance you select.
Top-of-mind questions for buyers
One host hour is one hour that a single inference instance runs, regardless of how many requests you process. Billing counts wall-clock time the instance stays active. Since the model is free, you pay only for that running compute time on the instance type you select.
Batch mode runs on ml.g5 instances for offline processing of collected inputs. Real-time mode runs on ml.g6e, ml.p4d, and ml.p5 instances for live, on-demand inference. Both bill per host hour. Choose batch for scheduled bulk jobs and real-time for interactive workloads.
Host hour billing counts the time the instance stays running, not the number of requests. An idle but active instance still accrues host hour charges. To stop charges, you must shut down the instance so no host hours accumulate.
www.ibm.com
Helpful?
Vendor refund policy
This model is provided by IBM completely free of charge. No payment is required to use it. Therefore, there are no purchases to refund.
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
IBM Granite Vision 3.3 2B is now available under the Apache 2.0 license for both research and commercial use. This release introduces advanced capabilities including image segmentation, structured doctag generation, and support for multi-page document processing. Built on an efficient architecture combining the Granite 3.3 decoder and SIGLIP2 vision encoder, it achieves top-tier accuracy on OCR and visual question answering benchmarks. Granite Vision 3.3 is ideal for automating document understanding in enterprise workflows across finance, healthcare, legal, and public sector domains.
Additional details
Inputs
- Summary
The model can be invoked by passing a prompt. Please see the sample notebook for details.
- Input MIME type
- application/json
https://github.com/ibm-granite-community/SageMaker/blob/main/granite-vision-3.3-2b/real_time_sample_input_data.json
https://github.com/ibm-granite-community/SageMaker/blob/main/granite-vision-3.3-2b/batch_sample_input_data.json
Input data descriptions
The following table describes supported input data fields for real-time inference and batch transform.
Field name | Description | Constraints | Required |
|---|---|---|---|
messages | This is the standard openAI body parameter included in the request body as part of the JSON payload for Chat models. | - | Yes |
Resources
Support
Vendor support
Support is not provided for this product.
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
IBM QRadar SIEM empowers security analysts and security operations teams with the visibility, automation and insights needed to quickly detect anomalies and uncover advanced threats in real-time.
IBM Verify Identity Access helps you simplify your users' access while more securely adopting web, mobile and cloud technologies.
IBM Granite 3.2 Instruct is an open-source model with controllable reasoning, offering strong performance and enhanced complex thinking.
IBM's Granite 3.0 8B Instruct is an 8B-parameter AI model for enterprise use, excelling in multilingual and code tasks; Apache 2.0 licensed.
Safeguard critical, sensitive, or regulated data wherever it resides
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.