Granite 4.2 LLMs: How They're Built (huggingface.co)

IBM has unveiled Granite 4.2, a new family of dense, decoder-only reasoning large language models (LLMs) that come in three sizes: 3 billion, 8 billion, and 30 billion parameters. These models, post-trained from the Granite 4.1 base models, enhance reasoning capabilities by utilizing a five-phase training strategy that includes a significant jump in context length to 512K tokens. The models feature a unique thinking/non-thinking switch, allowing for different modes of operation based on task complexity, and incorporate multi-stage reinforcement learning (RL) techniques to effectively employ tools within real sandboxed environments. The release of Granite 4.2 is significant for the AI/ML community as it not only improves upon instruction-following capabilities established in previous Granite iterations but also introduces robust reasoning features that can lead to more intelligent and context-aware applications. With functionalities for native tool calling, agentic RL, and the ability to operate in various reasoning modes, the models push the boundaries of LLM capabilities, fostering new possibilities in both research and practical applications in areas like software engineering and complex decision-making. Overall, Granite 4.2 sets a new benchmark in the development of reasoning-oriented language models while being accessible under the Apache 2.0 license.