Granite 4.2 LLMs: How They're Built

Granite 4.2 is a family of reasoning-focused language models released in three sizes that can produce chains of thought before answering questions and operate in different reasoning modes depending on task complexity. The models were built through a five-phase pre-training process on approximately 15 trillion tokens, followed by supervised fine-tuning on reasoning and tool-use data, and then reinforcement learning that teaches the larger models to act as agents by calling tools and running code in real environments. The larger models additionally learn to operate as agents through agentic reinforcement learning, allowing them to perform tasks like editing code, using terminals, and searching the web. All three models support native tool calling and are available under the Apache 2.0 license.