IBM has announced the release of its latest open-weight language model family, Granite 4.2, featuring enhanced agentic capabilities and a flexible licensing model. The company is offering three model sizes — 3 billion, 8 billion, and 30 billion parameters — each trained on approximately 15 trillion tokens, with a context window extending up to 512,000 tokens. This expansion marks a significant step forward in open-source AI development, especially with the inclusion of self-directed tool use and code execution in the larger models.
Agentic RL Training for Autonomous Models
The standout feature of the Granite 4.2 family is its use of agentic reinforcement learning (RL) for training the larger models. This method enables the models to learn how to interact with tools and execute code autonomously, without explicit programming for each task. This advancement positions the models as more adaptable and capable in real-world applications, such as software development, research, and automation.
Open Access Under Apache 2.0
All models in the Granite 4.2 lineup are now available under the permissive Apache 2.0 license, making them accessible for both commercial and research use. This move aligns with IBM’s broader strategy to promote open AI development and collaboration, while also competing with other open-source giants like Meta and Google. By offering these models freely, IBM is aiming to accelerate innovation in the AI space and attract developers and institutions looking to experiment with advanced language models.
Conclusion
The release of Granite 4.2 underscores IBM’s commitment to advancing open AI while pushing the boundaries of what language models can do. With built-in agentic capabilities and a generous licensing model, the models are poised to become valuable tools for developers, researchers, and enterprises alike.
