IBM has taken a significant step in the field of business AI with the launch of Granite 3.2, the latest version of its family of large language models (LLM). This new generation of models is designed to offer advanced reasoning, vision and security capabilities, and is available under the Apache 2.0 license in Hugging Face, as well as other platforms such as IBM Watsonx.AI, Ollama, Replicate and LM Studio.
The model has been trained using the IBM Docling open source tool kit, which processed 85 million PDF files to generate 26 million pairs of synthetic questions and answers. These capacities allow VLM to manage complex workflows and handle large volumes of documents efficiently.
One of the most prominent features of Granite 3.2 is its improved reasoning capacity, known as ‘Chain of Thought’. This functionality is available in models of 2 billion and 8 billion parameters, and can be activated or deactivated as necessary to optimize efficiency. Thanks to this capacity, the model of 8 billion parameters has shown significant improvements in instructions monitoring, such as Arenahard and Alpaca Eval, without compromising security or performance in other areas.
IBM expands the family of granite models with a new multimodal and reasoning AI
“Granite 3.2 extends further the abilities of AI with new reasoning functions, and we are excited to explore its potential in the development of autonomous solutions,” says David Tan, Cto de Crushbank.
Granite 3.2: Multimodal
IBM has also introduced reduced size options for Granite Guardian security models, which are now 30% smaller without losing performance. These models incorporate a new functionality called “Verbalized Confidence”, which allows a more precise and nuanced risk assessment. IBM’s strategy to offer smaller and specialized AI models has proven effective in various tests, as evidenced by the high performance of the Granite 3.1 8B model in the Salesforce LLM Benchmark for CRM.
In addition to Granite 3.2, IBM has launched the new generation of Tinymixers (TTM) models, designed to make long -term forecasts with less than 10 million parameters. These models are powerful tools for the analysis of trends in finance, economy, demand prognosis in the supply and planning chain of seasonal inventories in the retail sector.
Sriram Raghavan, vice president of IBM AI Research, highlights the importance of efficiency and accessibility in the next AI era: “The latest advances in Granite, focused on open solutions, represent another step forward to make AI more accessible, profitable and valuable for modern companies.”
