- Entire suite of Granite 3.0 models and the updated time series models are available for download on HuggingFace under the permissive Apache 2.0 license.
IBM unveiled its latest artificial intelligence models—Granite 3.0— to harness the growing interest in generative AI technologies among businesses.
The decision to release these models as open-source reflects IBM’s commitment to accessibility and innovation, contrasting sharply with industry competitors like Microsoft, which opt to monetise access to their AI solutions.
The Granite 3.0 models are designed for commercial use within IBM’s Watsonx platform, along with a paid tool that facilitates the deployment and customisation of these models within enterprise data centres. Additionally, these models are compatible with Nvidia’s software stack, enhancing their applicability for business integration.
Novel alignment technique
The collaboration emphasises the technological prowess of Nvidia’s H100 graphics processing units, which have been instrumental in the training of these advanced models.
A hallmark of the Granite 3.0 initiative is its adherence to open-source principles under the permissive Apache 2.0 license. This enables enterprises and the broader AI community to leverage the models’ superior performance, flexibility, and autonomy.
Unlike many existing large language models, which predominantly rely on publicly available datasets, Granite enables businesses to harness their own untapped data.
By employing the novel alignment technique, InstructLab, introduced with RedHat, IBM posits that businesses can achieve task-specific performance comparable to more extensive models, but at significantly lower costs—estimated to be between three to twenty-three times less than those for large frontier models.
Training process
Moreover, the release of Granite 3.0 reinforces IBM’s dedication to transparency and safety within AI development. Accompanying the models are comprehensive technical reports detailing the training datasets, the filtering and cleansing processes undertaken, and the performance results in relation to academic and enterprise benchmarks.
Notably, the Granite 3.0 8B Instruct model demonstrates superior performance across various metrics, notably eclipsing similar-sized open-source models from Meta and Mistral on Hugging Face’s OpenLLM Leaderboard and IBM’s own AttaQ safety benchmark.
The training foundation for the Granite 3.0 models is remarkable, encompassing over 12 trillion tokens across 12 natural languages and 116 programming languages.
The large-scale, two-stage training process is the result of extensive experimentation aimed at optimising data quality and selection.
Furthermore, anticipated enhancements by the year’s end include support for extended 128K context windows and advanced multi-modal document understanding capabilities.
Trained on more data
In addition to the Granite 3.0 models, IBM has announced advanced versions of its Granite Time Series models, which are now trained on three times more data and outperform significantly larger models from leading tech firms in key benchmarks.
The upgrade not only enhances modeling flexibility but also incorporates external variables and rolling forecasts, adding valuable capabilities for enterprise applications.
The entire suite of Granite 3.0 models and the updated time series models are available for download on HuggingFace under the permissive Apache 2.0 license.
The instruct variants of the new Granite 3.0 8B and 2B language models and the Granite Guardian 3.0 8B and 2Bmodels are available today 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.
To help provide developer choice and ease of use and support local, edge deployments, a curated set of the Granite 3.0 models are also available on Ollama and Replicate.