IBM's Granite 4.2 Adds Reasoning to Open LLM Line

IBM has published Granite 4.2, the newest iteration of its open large language model family, alongside a detailed look at how the models are built and trained. The release continues IBM's practice of shipping under the permissive Apache 2.0 license, which lets developers and companies use, modify, and deploy the models with minimal restrictions (Hugging Face).

What stands out in this version is the emphasis on reasoning. Granite 4.2 is positioned not just as a text generator but as a model tuned for more deliberate, multi-step problem solving—capabilities that have become table stakes as open models chase the frontier set by larger proprietary systems.

Why it matters

IBM's Granite line has always been aimed squarely at the enterprise, where licensing clarity and transparency about training often matter as much as raw benchmark scores. By documenting the architecture and training approach behind Granite 4.2, IBM is making a bet that buyers want to understand what's under the hood before they deploy.

• Open license: Apache 2.0, suitable for commercial use and fine-tuning.
• Reasoning focus: the family targets structured, multi-step tasks in addition to standard text generation.
• Transparency: IBM has published details on how the models were constructed and trained.

For teams weighing open alternatives to closed reasoning models, Granite 4.2 adds another credible option with an enterprise pedigree. The full technical write-up is available on IBM's Hugging Face blog.