IBM has released Granite 4.2, the newest generation of its open-weight large language models (LLMs) built for downloading, customization, and self-hosting. The Granite 4.2 lineup includes models with 3 billion, 8 billion, and 30 billion parameters.
Like earlier Granite models, Granite 4.2 uses a decoder-only architecture and includes a native 128,000-token context window. The 8B and 30B models also feature agent reinforcement learning, helping them perform tasks such as terminal commands, web searches, and interactions with external tools. While the 3B model supports tool use, it was not trained with the same level of specialized agent capabilities.
The most notable aspect of this release is its focus on inference. IBM describes Granite 4.2 as an “inference-focused release” designed to improve how the models handle complex, multistep tasks.
In the artificial intelligence industry, inference does not mean that an AI model understands a problem in the same conscious way a person does. Instead, it refers to functional reasoning, often supported by chains of thought. In this process, the model works through intermediate steps and uses those results to produce a final answer or complete an action.
Source: arstechnica.com