The company bets on more efficient, accessible, and optimized models for corporate environments
ARMONK, N.Y., February 27, 2025 – IBM has announced the arrival of Granite 3.2, the latest update to its collection of enterprise artificial intelligence models. With enhancements in reasoning, vision, and security, the company aims to offer more accessible and efficient tools without the high computational costs associated with larger models.
This new version is part of IBM’s focus on AI optimized for businesses, featuring more compact, scalable models under the Apache 2.0 license. Granite 3.2 is available on Hugging Face, with planned integrations in IBM watsonx.ai, Ollama, Replicate, and RHEL AI 1.5.
Key innovations of Granite 3.2
Advanced document processing with multimodal AI
Granite 3.2 includes a visual language model (VLM) that enhances the understanding of business documents. Its performance in tasks such as OCR, graph analysis, and information extraction is comparable to or exceeds that of larger models like Llama 3.2 11B and Pixtral 12B.
IBM has optimized this model using its Docling tool, processing 85 million PDF files and generating 26 million synthetic question and answer pairs, improving its accuracy in high-volume document environments.
Enhanced reasoning capabilities with “Chain of Thought”
Granite 3.2 introduces advanced reasoning options in its 2B and 8B models. This functionality can be toggled on or off as needed, allowing for resource consumption optimization.
- Higher accuracy: Double-digit increases in evaluations of instruction-following tests such as ArenaHard and Alpaca Eval.
- Scalable inference: Adjustments in the 8B model allow it to match the performance of larger models like Claude 3.5-Sonnet or GPT-4o in mathematical reasoning tests.
Improvements in security and efficiency
IBM has reduced the size of the Granite Guardian models by 30%, maintaining their effectiveness in risk detection. Additionally, a “verbalized confidence” feature has been added, allowing for more precise interpretations in threat assessment.
Granite 3.2 in the enterprise ecosystem
IBM’s strategy to develop smaller, optimized AI models continues to gain recognition in the sector. Granite 3.1 8B has achieved high ratings in tests such as the Salesforce LLM Benchmark for CRM, solidifying its effectiveness in enterprise applications.
David Tan, CTO of CrushBank, highlighted the impact of Granite models in his company:
“IBM has struck a balance between performance and scalability with its AI models. Granite 3.2 further expands these capabilities with advanced reasoning features, opening up new opportunities in intelligent automation.”
New models for long-term data forecasting
IBM has also introduced TinyTimeMixers (TTM), a new series of models specialized in time series analysis.
- Fewer than 10 million parameters.
- Capability of predicting trends up to two years.
- Applications in financial planning, logistics, and retail.
These models aim to provide more precise tools for data-driven decision-making.
A step closer to more accessible and efficient AI
IBM continues to invest in artificial intelligence that is more accessible and optimized to meet business needs.
“The AI of the future must be efficient and adaptable, allowing companies to extract maximum value without excessive costs. Granite 3.2 represents a significant advancement in that direction,” stated Sriram Raghavan, VP of IBM AI Research.
With Granite 3.2 and the TinyTimeMixers models, IBM reinforces its commitment to a more efficient, accessible AI aligned with the demands of the business sector.
Source: Artificial Intelligence News