Granite 4.2 LLMs: How They're Built

Granite 4.2 is a family of dense, decoder-only reasoning LLMs released in 3B, 8B, and 30B parameter sizes. Each model is pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy that extends the context window to 512K tokens. The models are then supervised fine-tuned on reasoning and agentic-trajectory data before undergoing a multi-stage reinforcement learning pipeline. The larger 8B and 30B models receive additional agentic RL training to learn tool use within sandboxed environments. All models support native tool calling and a "thinking" mode to produce a chain of thought before an answer.