News
ARMONK, N.Y. — IBM is introducing the latest version of its foundation AI model.
IBM released its open-source Granite 3.0 AI model family for enterprise applications, according to the company today.
The general-purpose Granite 3.0 8B and 2B large language models (LLMs) are intended to be compact and versatile models that can be fine-tuned with enterprise data and integrated in business environments.
The two models are designed to be "workhorse" models for enterprises tasks, such as retrieval-augmented generation (RAG), classification, summarization, entity extraction and tool use.
The Granite 3.0 models were trained on over 12 trillion tokens of data across 12 languages and 116 programming languages as well as results from several thousand experiments to optimize data quality, data selection and training parameters.
IBM reports the models can "outperform" or "match" similarly sized models on select academic and industry benchmarks.
The family's Granite Guardian models are designed for responsible AI and allow application developers to implement safety guardrails by checking user prompts and LLM responses for a variety of risks.
The line's Granite 3.0 Mixture-of-Experts models are set up to provide efficient inference and low latency, such as for CPU-based deployments and edge computing.
IBM's Granite 3.0 models and are available at HuggingFace under the Apache License 2.0.
