IBM Introduces Advanced Suite Of Open-Source AI Models: Granite 3.0

IBM Corp. has announced the release of its most advanced family of artificial intelligence models, Granite 3.0. The models will be made available as open-source under the permissive Apache 2.0 license.

The third-generation Granite flagship language models include those for general purpose/language (Granite 3.0 8B Instruct, Granite 3.0 2B Instruct, Granite 3.0 8B Base, Granite 3.0 2B Base), guardrails and safety (Granite Guardian 3.0 8B, Granite Guardian 3.0 2B) and mixed (Granite 3.0 3B-A800M Instruct, Granite 3.0 1B-A400M Instruct, Granite 3.0 3B-A800M Base, Granite 3.0 1B-A400M Base).

According to the company, the Granite 3.0 models were trained on over 12 trillion tokens on data taken from 12 different natural languages, and 116 different programming languages. By the end of the year, the 3.0 8B and 2B language models are expected to include support for an extended 128K context window and multi-modal document understanding capabilities.

With regard to data safety, IBM provides an IP indemnity for all Granite models on watsonx.ai so enterprise clients can be more confident of merging their data with the models.

Granite 3.0 8B and 2B are designed as 'workhorse' models for enterprise AI, for tasks such as retrieval augmented generation, classification, summarisation, entity extraction and tool use.

The Guardian models offer a comprehensive set of risk and harm detection capabilities, permitting application developers to implement safety guardrails by checking user prompts and LLM responses for various risks. Along with parameters such as social bias, hate, toxicity, profanity, violence and jailbreaking, these models provide RAG-specific checks such as groundedness, context relevance and answer relevance. 

The mixed models, Granite 3.0 1B-A400M and Granite 3.0 3B-A800M, are smaller and lightweight, and could be deployed for low-latency applications as well as CPU-based deployments.

IBM has also announced an updated release of its Granite Time Series models, which are pre-trained for multivariate time-series forecasting. The company said they are trained on more data and provide greater modelling flexibility.

The suite of Granite 3.0 models and the updated Time Series models are available on HuggingFace under the permissive Apache 2.0 license. The instruct variants of the Granite 3.0 8B and 2B language models and the Granite Guardian 3.0 8B and 2Bmodels are available for commercial use on IBM's watsonx platform. A selection of the Granite 3.0 models will also be available as Nvidia NIM microservices, and through Google Cloud's Vertex AI Model Garden integrations with HuggingFace, IBM said.