IBM Unveils Reasoning-Focused Granite LLMs

IBM has released its latest family of open-weight Granite large language models (LLMs) with 3 billion, 8 billion, and 30 billion parameters. The company's approach to model building differs from some competitors by focusing on dense, decoder-only reasoning models pre-trained from scratch.

The new Granite 4.2 family includes a 'reasoning-focused release' that allows models to run in thinking, non-thinking modes, or a low-effort mode for answering easy questions. Unlike other models of its class, Granite is text-only, although IBM offers related models such as Granite Vision 4.1.

The models were pre-trained on 15 trillion tokens and in five phases, including a long-context training phase that brings the family's context window to 512,000 tokens. The training set also included 1 trillion tokens of synthetic code generated by IBM's CodeAlchemy pipeline. For complex tasks, IBM used an agentic reinforcement learning step and RLHF alignment.