IBM open sources Granite 4.2 reasoning models with full training recipe

krvarshney · x · 2026-08-28

IBM released the Granite 4.2 reasoning model family (3B, 8B, 30B) under Apache 2.0, including the full training recipe—data mixture proportions, hyperparameters, RL environments, and chat templates. The models were pre-trained on 15T tokens with 512K context and post-trained via multi-stage RL in real software engineering, terminal, and web search environments. They feature three modes: thinking (default), non-thinking, and a low-effort mode.

Original post →

More from Models

Models channel →