IBM launches Granite 4.2 LLM family, highlighting Thinking Mode and agentic capabilities
IBM introduces Granite 4.2, a new family of large language models focused on reasoning and agentic workflows, released under the Apache 2.0 license.
On August 25, 2026, IBM launched Granite 4.2, a family of large language models (LLMs) engineered for advanced reasoning, available in three sizes: 3B, 8B, and 30B parameters.
A key highlight of Granite 4.2 is its support for "Thinking Mode," which enables Chain-of-Thought reasoning output wrapped in `<think>` tags. Users can toggle between modes as needed, including thinking, non-thinking, or low-effort modes. Additionally, the 8B and 30B models were trained using Reinforcement Learning (RL), unlocking agentic capabilities to use tools, write code, and search for information in live environments.
Under the hood, the model family was pre-trained on a 15-trillion-token (15T) dataset, supports a context window of up to 512K tokens, and was fine-tuned on over 7.2 million reasoning and agentic-trajectory examples. It is distributed under the Apache 2.0 license, allowing the developer community to freely build and innovate on top of it.
Granite 4.2 expands open-source options for Thai businesses and developers, offering access to models with stronger analytical reasoning and autonomous agent capabilities under a highly permissive license.