IBM Granite 4.2 Revolutionizes Open-Weights with Reasoning-Native Models

IBM has released Granite 4.2, a family of language models that marks a significant shift in the open-weights race. The three dense models, with parameter sizes of 3 billion, 8 billion, and 30 billion, are reasoning-native, meaning they can emit a chain of thought before answering.

The training pipeline for Granite 4.2 is also noteworthy. While earlier versions focused on instruction-following, the new models use multi-stage reinforcement learning to learn tasks such as code editing, terminal driving, and web searches in real sandboxes. The 8 billion and 30 billion parameter sizes are trained using agentic RL, which allows them to act in a more autonomous way.

The release of Granite 4.2 is part of a broader trend in the open-weights field, where models are shifting from competing on chat quality to reliable tool use, software engineering, and reasoning that holds up under real workloads. IBM's argument with this release is that a dense, well-trained, openly licensed model can join this conversation without resorting to mixture-of-experts scaling.