IBM has unveiled its latest and most advanced series of AI models, named Granite 3.0, during its TechXchange event.
The Granite 3.0 collection features several models, each tailored for different applications:
- General Purpose/Language: Available in 8B and 2B variants, with Instruct and Base configurations.
- Safety: Guardian models, also in 8B and 2B, are focused on safety guardrails.
- Mixture-of-Experts: A specialized set optimized for various deployment needs.
IBM asserts that the 8B and 2B language models from Granite 3.0 are on par with, if not superior to, similarly sized models from other leading providers. These models are built for enterprise-level tasks such as Retrieval Augmented Generation (RAG), summarization, classification, and entity extraction.
A distinguishing factor of Granite 3.0 is IBM’s dedication to open-source AI. Released under the Apache 2.0 license, the models offer high performance, flexibility, and control to enterprises and the wider AI community alike.
According to IBM, pairing their smaller Granite models with proprietary enterprise data, using their unique InstructLab alignment technique, enables organizations to achieve performance comparable to larger models while significantly reducing costs—by up to 23 times in some cases.
IBM continues to emphasize transparency and safety in its AI approach. The company has published a detailed technical report and a responsible use guide for Granite 3.0, outlining the datasets, data processing methods, and benchmark results. Moreover, IBM offers intellectual property indemnity for the models on its watsonx.ai platform, providing additional peace of mind for enterprise users.
Notably, the 8B Instruct model in the Granite lineup has outperformed similarly sized models from Meta and Mistral in academic benchmarks and leads in safety according to IBM’s AttaQ safety benchmark.
IBM is also rolling out the Granite Guardian 3.0 models, designed to provide safety guardrails by assessing user prompts and language model outputs for potential risks. These models include unique detection features for RAG-related issues like groundedness and relevance.
All Granite 3.0 models are available for download via HuggingFace, with commercial access options through IBM’s watsonx platform. IBM has also partnered with various ecosystem players to integrate Granite models into a broader range of enterprise solutions.
Looking ahead, IBM plans to enhance its AI portfolio further by developing more autonomous AI agents capable of solving complex problems. New features for AI agents are expected in IBM watsonx Orchestrate, along with expanded agent capabilities across its entire portfolio in 2025.
