IBM Debuts NextGen Granite 3.2 LLM Model, Powered By New Multi-Modal, Reasoning AI Capabilities

In an effort to deliver small, efficient, and practical artificial intelligence for enterprises, IBM Corp. has debuted the next generation of its Granite large language model family: Granite 3.2.

Alongside Granite 3.2 instruct, vision, and guardrail models, IBM also released the next generation of its TinyTimeMixers models (sub 10M parameters), with capabilities for longer-term forecasting up to two years into the future. These tools can aid long-term trend analysis, including finance and economics, supply chain demand forecasting and seasonal inventory planning in retail.

“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."

According to IBM, the new vision language model for document understanding tasks demonstrates performance that matches or exceeds that of significantly larger models—including Llama 3.2 11B and Pixtral 12B—on the essential enterprise benchmarks DocVQA, ChartQA, AI2D and OCRBench. In addition to training data, IBM used its own open-source Docling toolkit to process 85 million PDFs and generated 26 million synthetic question-answer pairs to enhance the VLM's ability to handle complex document-heavy workflows.

Granite 3.2 has chain of thought capabilities for enhanced reasoning in the 3.2 2B and 8B models, with the ability to switch reasoning on or off to help optimise efficiency. With this capability, the 8B model achieves double-digit improvements from its predecessor in instruction-following benchmarks like ArenaHard and Alpaca Eval without degradation of safety or performance elsewhere, IBM said.

Furthermore, with the use of novel inference scaling methods, the Granite 3.2 8B model can be calibrated to rival the performance of much larger models like Claude 3.5 Sonnet or GPT-4o on math reasoning benchmarks such as AIME2024 and MATH500.

The slimmed-down size options for Granite Guardian safety models maintain performance of previous Granite 3.1 Guardian models at 30% reduction in size. The 3.2 models also introduce a new feature called verbalised confidence, which offers risk assessment that acknowledges ambiguity in safety monitoring.

All Granite 3.2 models are available under the permissive Apache 2.0 license on Hugging Face. Select models are available today on IBM watsonx.ai, Ollama, Replicate, and LM Studio, and expected soon in RHEL AI 1.5.