At IBM’s annual TechXchange event, the company unveiled its latest and most advanced family of AI models, Granite 3.0. This third-generation flagship language model line is designed to outperform or match similarly sized models from leading providers on various academic and industry benchmarks, showcasing strong performance, transparency, and safety.
IBM’s commitment to open-source AI is evident in the Granite models, which are released under the permissive Apache 2.0 license. This unique approach combines performance, flexibility, and autonomy, catering to enterprise clients and the broader community.
Granite 3.0 Models Overview
The Granite 3.0 family includes a variety of models tailored for different functions:
- General Purpose/Language Models: Granite 3.0 8B Instruct, Granite 3.0 2B Instruct, Granite 3.0 8B Base, and Granite 3.0 2B Base.
- Guardrails & Safety Models: Granite Guardian 3.0 8B and Granite Guardian 3.0 2B.
- Mixture-of-Experts Models: Granite 3.0 3B-A800M Instruct, Granite 3.0 1B-A400M Instruct, Granite 3.0 3B-A800M Base, and Granite 3.0 1B-A400M Base.
The new 8B and 2B models are engineered as ‘workhorse’ solutions for enterprise AI, excelling in tasks such as Retrieval Augmented Generation (RAG), classification, summarization, and tool use. These models are designed to be fine-tuned with enterprise data, enabling seamless integration across various business environments.
IBM highlights that while many large language models (LLMs) are trained on publicly available data, a significant amount of enterprise data remains untapped. By pairing a smaller Granite model with enterprise data—especially using the innovative InstructLab alignment technique—IBM believes businesses can achieve task-specific performance rivaling larger models, potentially at a cost reduction of 3x to 23x.
Safety and Transparency in AI
IBM continues to emphasize transparency and safety in AI development. The Granite 3.0 technical report and responsible use guide detail the datasets used for training, including filtering and cleansing procedures. Additionally, IBM provides IP indemnity for all Granite models on its watsonx.ai platform, instilling confidence in enterprises merging their data with these models.
The Granite 3.0 models have demonstrated promising performance in benchmarks, leading the way on the Hugging Face’s OpenLLM Leaderboard and excelling in IBM’s AttaQ safety benchmarks. The training process involved over 12 trillion tokens, utilizing data from 12 natural languages and 116 programming languages, through a novel two-stage training method.
Introducing Granite Guardian 3.0
IBM also launched a new family of Granite Guardian models, designed to assist developers in implementing safety guardrails. These models check user prompts and LLM responses for various risks, including social bias, hate speech, and toxicity. The Granite Guardian 3.0 8B model outperformed previous models from Meta in harm detection accuracy and demonstrated solid performance in hallucination detection.
Availability and Future Plans
All Granite 3.0 models, along with updated time series models trained on three times more data, are available for download on Hugging Face under the Apache 2.0 license. The models are also available on IBM’s watsonx platform for commercial use and through integrations with NVIDIA NIM microservices and Google Cloud’s Vertex AI Model Garden.
IBM is committed to advancing enterprise AI, from models and assistants to the tools needed to tune and deploy AI for unique business needs. The upcoming release of the next-generation watsonx Code Assistant, powered by Granite code models, will offer general-purpose coding assistance across multiple programming languages.
Additionally, IBM is expanding its AI-powered delivery platform, IBM Consulting Advantage, to enhance the capabilities of its consultants, making Granite 3.0 the default model in Consulting Advantage for better client value. To learn more about Granite and IBM’s AI for Business strategy, visit https://www.ibm.com/granite.