IBM’s approach to foundation models
IBM Research positions the Granite family as an answer to the specific needs of enterprise adoption of language models: traceability of training data, compliance with sectoral regulations, contained size for on-premise deployment. The first significant release is Granite Code in May 2024, a series of models dedicated to code generation distributed under the Apache 2.0 licence.
On 21 October 2024, IBM publishes Granite 3.0, the first complete general-purpose generation of the family, also released under Apache 2.0 without additional restrictions. The range includes dense models of 2B and 8B parameters and Mixture-of-Experts variants with 1B and 3B active parameter configurations, designed to reduce latency in distributed inference scenarios.
Training mix and governance
A distinctive trait of the Granite family is the public documentation of the training mix: IBM makes explicit the datasets used, the proportions and the filtering processes applied to remove non-compliant content. This transparency — aligned with the diligence requirements foreseen by regulations such as the EU AI Act — allows adopting companies to conduct due diligence on source data and respond to audit questions on the provenance of model capabilities.
The training process is accompanied by filters on safety, known copyrighted content and sensitive material, with data quality pipelines documented in public technical reports.
Enterprise positioning
Granite models are designed to be competitive in the compact model tier — the most relevant size point for enterprise deployments — rather than competing with large frontier models. The 2B-8B sizes enable efficient inference on single enterprise GPUs or, in targeted cases, on modern CPUs, while the Apache 2.0 licence guarantees full freedom to use, modify and integrate in commercial products. IBM also integrates Granite models into its own watsonx platform, but without binding their use to that ecosystem.
Link: ibm.com/granite