IBM launches open-weight Granite 4.2 reasoning models

The latest Granite release is built to move out of the cloud and onto a developer’s own machine or infrastructure. IBM has launched Granite 4.2, an open-weight family designed for download and self-hosting, in 3B, 8B and 30B parameter versions. The new models are presented as the reasoning-focused chapter of IBM’s Granite language-model family.

The three variants share a native 128,000-token context window and a decoder-only architecture. That gives each model the same stated long-context capacity, while their training differs at the larger end of the range. The 8B and 30B models went through an agentic reinforcement learning stage aimed at expanded tool use: operating a terminal, searching the web and working with external tools.

The smallest model is not simply excluded from that workflow. The 3B version also supports tools, Ars Technica reports, but it did not receive the same level of specialized training as the 8B and 30B releases. The available report does not include comparative performance results for those tasks.

“Reasoning” also needs a narrow definition. In this context, it means functional reasoning: using chain-of-thought and carrying intermediate results through multiple steps. It does not mean that the models consciously understand a problem in the way a person does, a distinction Ars Technica makes explicit.

Concretely, Granite 4.2 gives teams a new set of locally runnable models to test and connect to tools, with the largest versions receiving targeted training for those interactions. The immediate benefit is the ability to run the models locally.