IBM Granite 4.2 with the full training recipe... IBM just published a reasoning model family under Apache 2.0 with the full training recipe. Data mixture proportions, per stage hyperparameters, RL environments, chat template modes. Everyone else is releasing weights and calling it open. Granite 4.2 comes in 3B, 8B, and 30B, pre trained from scratch on 15 trillion tokens with a five phase schedule out to 512K context, then post trained with multi stage RL. The 8B and 30B went through RL inside real software engineering, terminal, and web search environments. Every model exposes three modes through the chat template: thinking (default), non thinking, and a low effort setting with a short reasoning budget. The bet: enterprises will pay a premium for a reasoning model they can actually reproduce, audit, and continue training. Not just "weights available" but "we handed you the manual." For regulated industries and sovereign deployments, that distinction is worth more than a benchmark point or two. The unresolved question: which is the harder line for CIOs to cross in 2027, buying a closed reasoning model from a frontier lab, or standing up the internal MLOps to actually take advantage of a fully documented open one?