IBM’s Granite 4.2 Adds Native Reasoning, Agentic RL Training
IBM has released Granite 4.2, the latest version of its open-weight Granite model family aimed at enterprise use cases, according to Marktechpost. The update introduces native reasoning capabilities directly into the models, moving away from reliance on external prompting techniques to elicit step-by-step problem solving.
Marktechpost reports that IBM trained Granite 4.2 using agentic reinforcement learning, a method designed to improve how the models plan and execute multi-step tasks that resemble the workflows of autonomous AI agents. This approach targets enterprise scenarios where models must chain together actions, use tools, or make sequential decisions rather than simply generating single-turn text responses.
The release continues IBM’s strategy of offering open enterprise models through the Granite line, positioning the family as an alternative to closed frontier systems for businesses that want more control over deployment and customization. By combining native reasoning with agentic RL training, IBM is targeting use cases such as automated workflows, task orchestration, and decision support systems that require models to operate with greater autonomy.
Marktechpost’s coverage frames Granite 4.2 as part of a broader industry trend in which model developers are building reasoning and agentic capabilities directly into base models rather than treating them as add-on features accessed through specialized prompting or separate fine-tuned variants. IBM has positioned the Granite series as a foundation for enterprise AI applications since earlier releases in the line, and the 4.2 update extends that effort with a focus on tasks that go beyond conversational responses.
Further technical specifics, including benchmark results and model sizes, were not detailed in the available source material.
Based on reporting by www.marktechpost.com.
