Granite 4.2 adds reasoning at several sizes

IBM published a technical account of Granite 4.2, a reasoning-model family in 3B, 8B, and 30B sizes under Apache 2.0. The models include thinking and non-thinking modes, with agentic reinforcement learning described for the larger variants.

Context

Adjustable deliberation can help match the model to the operation. A simple extraction may not need the same budget as a complex analysis. The practical question is how often the additional reasoning changes correctness on representative work. Model size and thinking mode should both be included in a deployment comparison.