IBM's Granite 4.2: reasoning LLMs at 3B, 8B, and 30B, trained to think with tools
Granite 4.2 is IBM's first reasoning-native LLM family, three dense decoder-only sizes pretrained from scratch on ~15T tokens with a 512K context window. A five-phase pipeline adds chain-of-thought SFT, then multi-stage RL, including agentic RL where the 8B and 30B models learn tool use inside real sandboxed environments. Every model ships a thinking/non-thinking switch and native tool calling, all under Apache 2.0.