IBM Launches Enterprise Multimodal Inference Granite AI Models

**IBM Leads Practicality and Efficiency in Enterprise AI: Unveiling Granite 3.2 and TinyTimeMixers** IBM has announced the dawn of a practical and cost-effective AI era with the launch of 'Granite 3.2,' the next-generation core of its enterprise large language model (LLM) portfolio. This new product is broadly available on major platforms including Hugging Face (Apache 2.0 license), IBM watsonx.ai, and Ollama, and will soon be accessible on Red Hat Enterprise Linux (RHEL) AI 1.5. A key innovation of Granite 3.2 is its new **Vision Language Model (VLM)**, specialized for interpreting complex documents. Trained on over 85 million PDFs and 26 million synthetic question-answer pairs, this model has demonstrated excellent performance, on par with or surpassing existing large models like Llama 3.2 11B and PixTral 12B, across essential enterprise benchmarks such as DocVQA. Additionally, the 3.2 2B and 8B models introduce **enhanced 'Chain of Thought'** capabilities, allowing users to activate or deactivate reasoning functions as needed to maximize computing resource efficiency. Notably, the 8B model achieved double-digit performance improvements on ArenaHard and AlpacaEval benchmarks with this dynamic reasoning approach, and it can be fine-tuned to rival models like Claude 3.5-Sonnet and GPT-4o on mathematical reasoning benchmarks such as AIME2024 and MATH500, offering both powerful performance and flexibility. This is central to IBM's 'practical AI' strategy, which reduces unnecessary computational load for simpler tasks. There have also been advancements in safety. The **Granite Guardian safety model** has increased efficiency by reducing its size by 30% while maintaining existing performance, and through a new feature called 'verbalized confidence,' it recognizes ambiguities in risk assessment and supports more sophisticated safety monitoring. Alongside Granite 3.2, IBM also unveiled **'TinyTimeMixers (TTM),'** a next-generation time series model capable of long-term predictions up to two years into the future. Operating with fewer than 10 million parameters, TTM is expected to bring innovation to long-term business forecasting, including financial and economic trends, supply chain demand, and retail inventory planning. Sriram Raghavan, Vice President of IBM AI Research, emphasized, "The next era of AI will be defined by efficiency, integrability, and practicality," stating that Granite 3.2 will contribute to enhancing enterprises' AI accessibility, cost-effectiveness, and the value of AI itself through open solutions. IBM is supporting the creation of new business value by providing critical AI solutions to companies that seek powerful performance without excessive computing costs.