Trending Articles
Trending Topics
Trending
IBM has announced a significant expansion of its Granite 3.2 large language model family with the introduction of three small and efficient AI models.
Together, they promise vision, forecasting and experimental reasoning capabilities.
The new vision language model (VLM), IBM Granite Vision 3.2 2B, focuses on document understanding. In this area, IBM claims, it outperforms the much larger models of Llama 3.2 11B and Pixtral 12B.
IBM says it achieved this by pairing its own training data with 85 million PDFs. It then “generated 26 million synthetic question-answer pairs to enhance the VLM’s ability to handle complex document-heavy workflows”, to quote the press release.
All models have been trained on curated and enterprise-specific data, ensuring their relevance and effectiveness in real-world business scenarios. As opposed to the likes of OpenAI’s GPT4o and similar non-enterprise LLMs, which are often based on more generic data.
The release highlights IBM’s strategy to deliver smaller, specialised AI models for enterprises, echoing an article by Juan Bernabe Moreno, Director of IBM Research Europe in the UK & Ireland. And it believes this strategy is working, citing the the Granite 3.1 8B model’s performance in the Salesforce LLM Benchmark for CRM.
Digging into these results shows that it performed well for accuracy and speed, mid-table for cost but there is still work to be done in the trust & safety area. To quote Salesforce: “This is a critical dimension for most businesses, and includes privacy, safety, and general truthfulness. Model fine-tuning and prompt engineering can improve these scores.”
Despite the low scores in the Salesforce test, IBM is keen to emphasise that its approach to AI development is based on trust and transparency.
That’s nothing new. But Granite 3.2’s release marks a shift towards smaller, more efficient models from the company.
“The next era of AI is about efficiency, integration, and real-world impact – where enterprises can achieve powerful outcomes without excessive spend on compute,” said Sriram Raghavan, VP, IBM AI Research. “IBM’s latest Granite developments focus on open solutions demonstrate another step forward in making AI more accessible, cost-effective, and valuable for modern enterprises.”
CrushBank CTO David Tan said that it has already seen IBM’s AI models “deliver real value for enterprise AI”. He added: “Granite 3.2 takes it further with new reasoning capabilities, and we’re excited to explore them in building new agentic solutions.”
Notably, IBM is also releasing the next generation of its TinyTimeMixers (TTM) models, which have less than 10 million parameters. It claims these can forecast up to two years into the future. “These make for powerful tools in long-term trend analysis, including finance and economics trends, supply chain demand forecasting and seasonal inventory planning in retail,” said the press release.
It is available under the Apache 2.0 licence on Hugging Face, as with previous Granite models. The family is also now available on IBM’s own watsonx.ai platform, Ollama, Replicate and LM Studio. IBM says it is “expected soon” in RHEL AI 1.5.
