At the annual TechXchange event, IBM announced the release of Granite 3.0, a family of artificial intelligence models licensed under Apache 2.0, emphasizing the company's commitment to open-source AI.
The new Granite 8B and 2B models are designed for enterprise AI applications, offering advanced performance at a lower cost for tasks such as Retrieval Augmented Generation (RAG), classification, summarization, entity extraction, and tool utilization. These models can be trained with company data and seamlessly integrated into various business environments.
Unlike most large language models (LLMs) trained on public data, Granite leverages underutilized enterprise data. By combining Granite with business data and using retraining methods like InstructLab, introduced by IBM and RedHat, companies can achieve competitive performance at significantly lower costs.
The release of Granite 3.0 reinforces IBM's commitment to transparency, security, and reliability. Comprehensive technical documentation details the datasets used for training, filtering, cleaning, and curation processes, along with performance data against major academic and business benchmarks. IBM also offers intellectual property guarantees for all Granite models on watsonx.ai, ensuring clients can safely integrate their data.
The Granite 3.0 models have been trained on over 12 trillion tokens across 12 languages and 116 programming languages, utilizing a two-phase training method optimized through thousands of experiments. By the end of the year, the 8B and 2B models will support extended context lengths up to 128K and multimodal document understanding capabilities.
Additionally, IBM introduced the Granite Guardian model family, enabling application developers to implement security measures by controlling user prompts and LLM responses against various risks. The Granite Guardian 8B and 2B models provide comprehensive risk detection features and can be used alongside any open-source or proprietary AI model.
The complete suite of Granite 3.0 models and updated time series models is available for download on HuggingFace under the unrestricted Apache 2.0 license. Instruction variants of the new Granite 3.0 models and Granite Guardian models are now commercially available on the IBM watsonx platform. Selected Granite 3.0 models will also be accessible as Nvidia NIM microservices and through Google Cloud's Vertex AI Model Garden with HuggingFace. A curated selection of Granite 3.0 models is also available on Ollama and Replicate.
This latest generation of Granite models enhances IBM's open-source catalog with powerful, purpose-built LLMs. IBM has collaborated with partners such as AWS, Docker, Domo, Qualcomm AI Hub, Salesforce, SAP, and others to integrate Granite models into their offerings or make them available on their platforms.