IBM launched Granite 3.0 8B and 2B models under the Apache 2.0 license, new models designed for CPU-based deployments and edge computing and the next-generation of Watsonx code assistant. In addition, IBM said Granit models will be the default for Consulting Advantage, an AI delivery platform used by the company's consultants.
Big Blue announced the latest Granite large language models (LLMs) at its TechXchange event. IBM said the Granite family of models is under the fully permissive Apache 2.0 license for enterprise use cases.
"IBM keeps advancing its Granite model family, alleviating the concerns that it was simply being tossed over to open source in move that CxOs have seen too often. IBM has the knowhow and data to maintain these models. The future use of Granite models in next-gen apps looks brighter than ever," said Constellation Research analyst Holger Mueller.
IBM's Granite 3.0 family includes the following:
- Granite 3.0 8B-Instruct, Granite 3.0 2B-Instruct, Granite
- 3.0 8B Base, Granite 3.0 2B Base for general purpose and language use cases.
- Granite Guardian 3.0 8B, Granite Guardian 3.0 2B focused on guardrails and safety.
- Granite 3.0 3B A800M Instruct, Granite 3.0 1B A400M Instruct,
- Granite 3.0 3B A800M Base, Granite 3.0 1B A400M Base as mixture-of-experts models.
The Granite 8B and 2B models are designed to be workhorses that deliver strong performance and cost efficiency for RAG, summarization and classification. IBM expects these models to be adopted and then fine-tuned by businesses looking to avoid the costs associated with larger models. IBM discloses the data sets used to train Granite and provides IP indemnity on watsonx.ai.
- IBM open sources Granite models, integrates watsonx.governance with AWS SageMaker
- IBM Delivers AI How It Is Meant to Be With Watsonx
IBM also released benchmarks for Granite 8B.
According to IBM, the Granite mixture-of-experts models (A800M) are designed for low-latency environments, edge use cases and CPU-based inference deployments.
As for the Granite Guardian 3.0 models, IBM said the family is designed to check user prompts and LLM responses for various risks including bias, toxicity and jailbreaking.
Going forward, IBM said it will use the Granite models and extend them with AI agent capabilities for autonomy. Granite 8B features agentic capabilities for workflows. These capabilities will be rolled out in 2025 with prebuilt agents for use cases.
