IBM has recently released the Granite 3.2 series of open-source AI models, enhancing inference capabilities and introducing its first vision-language model (VLM) while continuing advancements in time-series models. Notably, a toggleable reasoning mode is available to help reduce unnecessary computing costs.
The Granite series includes language models, vision-language models, and specialized models for tasks such as coding, time-series analysis, geospatial data processing, and cybersecurity. In terms of language models, Granite 3.2 is available in 2B and 8B parameter versions, now with improved chain-of-thought (CoT) reasoning for more sophisticated problem-solving.
IBM emphasized that while CoT reasoning significantly enhances inference tasks, it also demands substantial computing resources and isn't always necessary. To optimize efficiency, Granite 3.2 models now feature a programmable toggle, allowing users to enable or disable reasoning modes based on their specific needs.
A key addition to the lineup is Granite Vision 3.2 2B, a compact vision-language model designed for enterprise applications. It supports both text and image inputs and is tailored for daily business operations.
IBM reports that Granite Vision 3.2 2B outperforms larger models, including 4B and 11B parameter variants, on enterprise document-related benchmarks such as DocVQA and ChartQA. Competitors in these tests include Microsoft Phi-3.5 and Meta Llama-3.2.
IBM also highlighted the popularity of its time-series models, Tiny Time Mixers (TTMs), which have been downloaded over eight million times on the Hugging Face platform.
Designed specifically for time-series forecasting, TTMs have low parameter counts (in the millions), making them highly efficient even on CPU-only devices. These models are well-suited for applications such as financial trend analysis, supply chain demand forecasting, and seasonal inventory planning in retail.
IBM AI Research Vice President Sriram Raghavan emphasized that the next phase of AI will focus on efficiency, integration, and real-world impact. He stressed that enterprises should aim to maximize AI benefits without excessive computational costs.
Beyond its own models, IBM's watsonx.ai platform also supports models from Meta Llama, Mistral AI, Google Flan-T5, and others, providing enterprises with a diverse set of AI solutions.
Article edited by Jack Wu
