IBM launched the latest generation product, Granite 3.2, of its Granite language modeling family at the end of February, continuously promoting small, efficient, and enterprise specific AI to create benefits for practical applications.

All Granite 3.2 models are licensed under the relaxed Apache 2.0 open source license and can be downloaded on Hugging Face. Some models are now available on IBM Watsonx.ai, Ollama, Replicate, and LM Studio, and are expected to soon support RHEL AI 1.5, injecting stronger AI capabilities into enterprises and open source communities.

Main highlights

• New Visual Language Model: Designed specifically for understanding document tasks, it outperforms even larger models such as Llama 3.2 11B and Pixral 12B in key enterprise benchmark tests DocVQA, ChartQA, AI2D, and OCRBench. In addition to powerful training data, IBM also utilizes its open-source Docling toolkit to process 85 million PDF files and generate 26 million synthetic question answer pairs, enhancing the visual language model's ability to handle large file workflows.

• Enhanced reasoning function: The 2B and 8B models of Granite 3.2 have added a "Chain of Thought" (CoT) reasoning mechanism, and users can turn on or off the reasoning function to optimize efficiency. Through this ability, the 8B model performs well in benchmark tests following instructions such as ArenaHard and Alpaca Eval, achieving double-digit superiority over the previous generation without affecting the safety or performance of other domains. In addition, through innovative reasoning extension methods, the Granite 3.2 8B model can be adjusted to approach the performance of Claude 3.5 Sonnet or GPT-4o on mathematical reasoning benchmarks such as AIME2024 and MATH500.

• The Granite Guardian security model is lighter: while maintaining the performance of the Granite 3.1 Guardian model, the model size is reduced by 30%. In addition, the Granite 3.2 series also introduces a new feature called Verbalized Confidence, which provides more refined risk assessment and helps security monitoring systems identify uncertainty.

IBM continues to promote enterprise specific small AI model strategies and has demonstrated high performance in testing. For example, the Granite 3.1 8B model achieved high scores in the Salesforce Large Language Model CRM benchmark test, demonstrating its accuracy and reliability in practical applications.


The IBM Granite model family has a vast ecosystem of partners, and many leading software companies have embedded Granite models into their technology. Granite 3.2 is an important advancement by IBM in promoting enterprise specific small AI, reflecting IBM's product strategy of providing small, efficient, and practical AI.


David Tan, Chief Technology Officer of CrushBank, said, "At CrushBank, we have witnessed firsthand how IBM's open and efficient artificial intelligence models bring real value to enterprise AI - achieving the right balance between performance, cost-effectiveness, and scalability. Granite 3.2 goes further with new inference capabilities, and we are pleased to explore these features when building new agent solutions. "


Granite 3.2 is an important step in IBM's product portfolio and strategic development, aimed at providing small and practical AI for enterprises. Although the thought chain performs strongly in reasoning tasks, it requires a significant amount of computing resources and not all tasks must be enabled. Therefore, IBM added a programmatic switch function to the Granite 3.2 model, allowing users to turn on or off inference mode according to their needs; The model can run simpler tasks without enabling inference to reduce unnecessary computational costs.


In addition, other inference techniques, such as inference scaling, have shown that the Granite 3.2 8B model can match or even surpass the performance of larger models in standard mathematical inference benchmark tests. Continuously developing this reasoning technology is also a key focus of the IBM research team to further enhance the efficiency and application scope of AI.


In addition to the instruction, visual, and protection models of Granite 3.2, IBM has also launched a new generation of TinyTimeMixers (TTM) time series models, which have parameters of less than 10 million and have long-term prediction capabilities, capable of making long-term predictions for up to two years. These models provide powerful tools for long-term trend analysis, suitable for financial and economic trend analysis, supply chain demand forecasting, and seasonal inventory planning in the retail industry.


Sriram Raghavan, Vice President of AI Research at IBM, stated that the next era of AI will focus on efficiency, integration, and the impact of practical applications - businesses should be able to achieve powerful AI benefits without excessive consumption of computing resources. IBM's latest Granite model development focuses on open solutions, gradually promoting the popularization of AI, making it more cost-effective, and creating greater value for modern enterprises. "