IBM has released the third generation of its most advanced family of AI models to date, the Granite language models, at its annual TechXchange event on Monday.

Released under the Apache 2.0 license, Granite 3.0 open-source models are designed for enterprise clients, delivering tasks including Retrieval Augmented Generation (RAG), classification, summarisation, entity extraction, and tool use.

IBM explained that many large language models (LLMs) are trained on publicly available data, leaving most enterprise data untapped.

The new Granite model is combined with enterprise data through InstructLab with the aim of enabling task-specific performance at reduced costs.

Introduced by IBM and RedHat in May, InstructLab is an open source project designed to lower the cost of fine-tuning LLMs by providing communities the tools to update them without having to retrain the model from scratch.

IBM said the Granite 3.0 models were trained on more than 12 trillion data tokens from 12 different natural languages and 116 different programming languages.

It added that by the end of the year, the 3.0 8B and 2B language models should include support for an extended 128K context window and multimodal document understanding capabilities.

Additionally, IBM has introduced an upgraded variant of its pre-trained Granite time-series models, the first versions of which were released earlier this year, aiming to offer higher performance.

According to IBM, these new models are able to outperform larger models including Google, Alibaba and others.

IBM also offers an intellectual property guarantee for all Granite models on Watsonx, aiming to offer more security and transparency to enterprises seeking to merge their data with the models.

Watsonx is IBM's commercial generative AI and scientific data platform based. The cloud-based platform, which supports multiple LLMs along with IBM’s Granite, offers a studio, data store, and governance toolkit.

---