With Granite 3.0, IBM aimed to provide businesses with a more efficient, effective, and accessible AI. Granite 3.2, launched on February 27, introduces major advancements in reasoning, security, and document comprehension, while staying true to IBM's philosophy: an open-source AI, optimized for professional use. Like its predecessors, the models are distributed under the Apache 2.0 license and can be downloaded on HuggingFace.
With Granite 3.2, IBM strengthens its position in the AI market by offering a range of models capable of competing with larger competitors while optimizing resource consumption.
This update introduces:
- An advanced Visual Language Model (VLM): the new multimodal model, Granite Vision 3.2 2B, trained to handle both image and text inputs, excels in document understanding and analysis, surpassing models like Llama 3.2 11B and Pixtral 12B on key benchmarks (DocVQA, ChartQA, AI2D, and OCRBench). IBM leveraged its Docling toolkit to process 85 million PDFs and generate 26 million question-answer pairs, thereby enhancing the model's robustness.
- Unprecedented flexibility in reasoning: The Granite 3.2 Instruct 2B and Instruct 8B models allow for the activation or deactivation of their reasoning capability ('chain of thought') to optimize their efficiency. Thanks to this innovation, the 8B model achieves double-digit improvements over its predecessor on benchmarks such as ArenaHard and Alpaca Eval, and competes with Claude 3.5-Sonnet and GPT-4o in mathematical reasoning (AIME2024 and MATH500).
- Enhanced security: The Granite Guardian 3.2 range, specifically designed to meet critical security and compliance needs of businesses, reduces model size by 30% while maintaining reliability. It introduces a new approach called verbalized trust, which refines risk assessment by considering areas of uncertainty.
An AI Better Suited to Business Needs
In parallel, IBM is launching a new generation of its TinyTimeMixers (TTM) models, which have less than 10 million parameters and allow for the analysis of financial and economic trends, inventory management planning, and supply chain optimization. The latest addition, TTM-R2.1, extends forecasts to one week.
Sriram Raghavan, VP, IBM AI Research, comments:
"The next era of AI is about efficiency, integration, and real-world impact - where businesses can achieve powerful outcomes without excessive computational costs. IBM's latest Granite developments, focused on open solutions, represent a new advancement to make AI more accessible, cost-effective, and valuable for modern businesses."
The Granite 3.2 models are available on IBM watsonx.ai, Ollama, Replicate, and LM Studio, and are expected to be available soon on RHEL AI 1.5.