IBM today introduced a new family of large language models called Granite 4.2. These models are offered in three sizes—3 billion, 8 billion, and 30 billion parameters—and have been trained entirely from scratch with about 15 trillion tokens. Granite 4.2 is built on a decoder‑only architecture and supports text windows of up to 512 thousand tokens.
A notable feature of these models is the “thinking” mode, which allows the model to reason in a chain before generating a response. In addition, the 8 billion and 30 billion models, after the pre‑training stage, undergo an agentic RL (reinforcement learning) block that gives them the ability to call external tools, execute code, work in a terminal, and search the web. All versions can call internal tools.
Granite 4.2 is released under the Apache 2.0 license and is optimized for use in cloud, on‑premises, and edge environments; therefore it can be used in enterprise, research, and commercial applications. IBM presents these models as an important step toward providing reliable and customizable AI for multi‑step and agentic tasks.